Professional Documents
Culture Documents
1
Proceedings of the 19th International Conference on VLSI Design (VLSID’06)
1063-9667/06 $20.00 © 2006 IEEE
INL and DNL requirement, Lin and Bult [2] have done calculated in Table 1. Line 4 shows the required analog
the area analysis for 10-bits resolution. In this paper, an area variation for DNL = 0.5 LSB with segmentation.
analysis for 16-bits resolution is done. This issue is The two extremes of this line are in accordance with
analyzed via a MATLAB simulation. First an analysis is the calculated values of area requirement for binary
done for only binary coded and only thermometer coded and thermometer coded architecture, as shown in Table
DAC. If same analog area is taken for both the 1. Line 5 represents the digital area required. Here it is
architecture, and if thermometer coded architecture gives assumed that digital area required per current cell is
DNL of σ, then binary weighted gives 2 N σ, where N approximately 0.35 times of required analog area for
that current cell. This assumption is taken in
is the total number of bits, and σ is standard deviation of
accordance with the available Lin & Bult Plot for 10-
current source value. Also, for this same area, both
bit resolution [2]. The bold curve, (line 6) in the plot
architectures give INL of 0.5* 2 N σ [2]. represents the total required area variation with
Since, area α (1/σ2), so a calculation is done for segmentation for INL = 1.0 LSB, and DNL = 0.5 LSB.
both the architectures, assuming that Aunit is the minimum
area required by the thermometer architecture to obtain
DNL = 0.5 LSB. The results are summarized in Table 1.
1
Table 1. Area requirements for specific INL/DNL 6 A
2
Requirement Thermometer Binary
Coded Weighted 3
2
Proceedings of the 19th International Conference on VLSI Design (VLSID’06)
1063-9667/06 $20.00 © 2006 IEEE
of the current sources is given a random value that has
been derived from a gaussian distribution with mean
value I and a standard deviation σI. For every digital code
the output current of the D/A converter is calculated and
compared to the ideal value. If the difference is larger
than an LSB, even for only one digital code, the D/A
converter is regarded as not functional and is therefore
rejected. This procedure is repeated for every σI , for 100 Yield
times to obtain reliable results. The yield is given by the (%)
ratio of the number of functional D/A converters (INL <
1.0 LSB and DNL < 0.5 LSB) to the total number of try-
outs. Fig 3 shows the yield variation with σI for 16-bit
segmented resolution for 100 Monte Carlo runs. This
curve shows that yield is 97 % for σI / I equal to 0.26 %.
This value of σI / I is used in the calculation of W and L
of current source transistor (discussed in section IV-B).
Using this value of σI / I, INL and DNL plots have been
plotted for 16-bit Resolution (13-bits Thermometer coded σI/I (%)
3-bits Binary weighted) from 100 Matlab simulations, as
shown in Fig 4 (a) and (b). From these curves, the
maximum INL and maximum DNL comes out to be 0.8 Fig 3. Yield variation with σI for 16-bit
LSB, and 0.03 LSB respectively, which satisfies the INL resolution (for 100 Monte Carlo simulations)
& DNL requirement.
4. Circuit Architecture & Results
0.8
0.2
The complete circuit with all the blocks is
shown in Fig 5. In this architecture, 13-bits of
INL 0
( in LSB) Thermometer coded have been splitted into two groups
-0.2 of 7-bits and 6-bits, to reduce the area. But this
-0.4 introduces a glitch when output jumps from 6-bit block
to 7-bit block. Fig 6 shows the circuit of one unit
-0.6
current cell. It consists of an analog part and a digital
-0.8 part. The analog part consists of a differential switch
-1.0 and a cascoded current source. The digital part consists
0 1 2 3 4 5 6 7
4
input code ( x 10 ) of a decoding logic and a latch. The decoding logic is
0.04 equivalent to an AND–OR gate function and the latch
is essential for timing synchronization, as all the
0.03
current cells should switch at the same time.
0.02
Basic architecture of this Decoding Logic
0.01 block is shown in Fig. 7(a). With this structure, there
DNL
(in LSB)
occurs one problem. When one column is completely
0
selected, and one more cell is to be added in the output
-0.01 current, then requirement is that the value of all ROW
should become ‘0’, and the value of next COL should
-0.02
become ‘1’, and these two actions should occur
-0.03 simultaneously, where ROW & COL are the outputs of
-0.04
row and column decoder. But if there is little bit delay,
0 1 2 3 4 5
4
6 7 then a glitch will occur. In the modified structure, some
input code ( x 10 )
delay is added to ROW and COL input, as compared to
COL+1 input (next adjacent COL value). Also, COL is
more delayed than ROW. This causes COL +1 to be ‘1’
Fig 4. (a) INL & (b) DNL Plots for σI / I = 0.26% earlier, which maintain the current status of the total
current, and then ROW input goes to ‘0’, so that output
3
Proceedings of the 19th International Conference on VLSI Design (VLSID’06)
1063-9667/06 $20.00 © 2006 IEEE
of (ROW ∩ COL) doesn’t give ‘1’, which removes the current cell, the current source transistor is cascoded by
problem of glitch Fig 7(b). If the switches interrupt the the transistor MCD.
current flow at any time point, then the current source B. Mismatch Analysis
charges the common source point of both the switches
towards VDD. And then current source and cascoding As defined by Pelgrom, mismatch “is the
transistor may enter into linear region, which may process that causes time-independent random
deteriorate the dynamic -performance considerable. variations in the physical quantities of identically
Therefore the controlling signals must not cross in the designed devices” [1]. Essentially, this means that
middle of the supply voltage. So, the PMOS transistor each current source in the matrix generates a current
size of both inverters (Inv_1 & Inv_3) should be larger that varies slightly from the desired current, IREF.
than the NMOS transistors to decrease the rise time. Therefore the current sources have to be designed in
Inverters (Inv_5 & Inv_6) are used between digital and such a way that random variations do not degrade the
analog part to change the analog value of logic ‘0’ at the performance of the circuit below its desired
input of switches, to increase the dynamic range of output specifications. Pelgrom’s paper has become the
voltage. standard for analysis of transistor matching, and thus
The critical transistor in the whole analog part is his formula for the standard deviation of saturation
the current source transistor (MC). A little bit change in current for two identically sized devices was used for
VT and β can cause a drastic change in the output current. the design. This formula is:
So, VT and β mismatch effects have been taken into the
consideration in the W & L calculation of current source σ 2 (Id ) 4σ 2 (VTo ) σ 2 ( β )
= + ,
(section IV-B). To increase the output impedance of the I d2 (VGS − VTo ) 2 β2
Inv 1 Inv 2 MC
COL s V1
ROW M1
COL+1 B1 s bar V2 MCD
Inv 5
SW1 SW2
Inv 6
out out bar
M2 Inv 4
Inv 3
COL s COL
ROW s
COL+1 ROW
s_bar COL+1
s bar
(a) (b)
Fig 7. Decoding Logic Structure
4
Proceedings of the 19th International Conference on VLSI Design (VLSID’06)
1063-9667/06 $20.00 © 2006 IEEE
400
400
300
300
200
200
Current
(µA) 100 Current
(µA) 100
0
0
-100
-100
-200
0 100 200 300 400 500 600 -200
30 40 50 60 70
Time (ns)
Time (ns)
(a) (b)
Fig 8. Glitch problem in 384 µA current cell during transition
300 300
200 200
Current Current
(µA) 100 (µA) 100
0 0
-100 -100
-200
0 100 200 300 400 500 -200
30 40 50 60 70
Time (ns) Time (ns)
(a) (b)
Iout
Iout -+ Fig 10. Glitch variation with Finger concept
49
48
Current
(mA) 47
46
0 10 20 30 40 50 60
Iout - Time
Time(µs)
(ns)
48.854 Iout - (a)
(a (a)
48.849
Current
(mA)
48.844
48.839
48.834
6
Proceedings of the 19th International Conference on VLSI Design (VLSID’06)
1063-9667/06 $20.00 © 2006 IEEE
A 1 .2-V, 600-MS/s, 2.4-mW DAC for WLAN
802.11 and 802.16 Wireless Transmitters
Nicola Ghittoril, Andrea Vignal, Piero Malcovati2 Stefano D'Amico3, Andrea Baschirotto3
'Dept. of Electronics, 2Dept. of Electrical Engineering, University of Pavia, Pavia, Italy
3Dept. of Innovation Engineering, University of Lecce, Lecce, Italy
Email: [nicola.ghittori,andrea.vigna,piero.malcovati] @unipv.it; [stefano.damico,andrea.baschirotto] @unile.it
Abstract- For the present and up-coming WLAN applications scaling will allow the use of higher-speed and lower-power
(802.11a/g, 802.11n, 802.16), a transmission baseband architec- (proportional to FSCV2D) digital part. On the other hand the
ture uses a 600-MS/s current-steering DAC with a passive output same trend implies analog sections with poorer performance,
load to perform the baseband signal processing, avoiding the use
of any active reconstruction filter. In a 0.13-tm CMOS technology due to the limited linearity achievable with the available supply
the DAC consumes 2.4 mW from a single 1.2-V supply voltage. voltage (which can be 1.2 V or lower).
The DAC exhibits a full-scale SFDR of 68 dB for an input As the new standards will present a larger signal bandwidth
signal frequency of 12 MHz and a full-scale dynamic range of (14 MHz for the upcoming WLAN 802.16 for instance), the
9.7 bits between 0 and 10 MHz. These data correspond to the use of traditional TX baseband architectures will result in
best reported Figure of Merit, if compared with state-of-the-art
digital-to-analog converters. a more and more critical design of the analog filters, since
their cut-off frequency has to be increased (with an increasing
I. INTRODUCTION sensitivity to the lower CMOS gain and to the non-dominant
Present baseband sections for telecom standards transmitters poles). Moreover they have to guarantee the required dynamic
consist of the cascade of a digital-to-analog converter (DAC) range and linearity at the lower supply voltage imposed by
receiving the digital signal processor (DSP) bit-stream and the scaled technologies. These points would determine an
an analog reconstruction (or smoothing) filter, which has to increasing power consumption of the analog filtering blocks.
suppress the DAC spectral images. A digital interpolator filter Thus, maintaining the ratio between the conversion fre-
is required between the DSP (which typically operates at quency and the signal bandwidth equal to about 10 would
Nyquist frequency) and the DAC to enhance the data-rate to result in a digital part underperforming its effective possibility,
the desired value. This architecture is shown in Fig. 1. The without exploiting the benefits of the reduced supply voltage,
design of this baseband section has to optimize the trade-off and in a power-consuming analog part introducing distortion
between two possible approaches: and noise. As a consequence, the scheme of Fig. 1 will be
. a low DAC conversion frequency, which implies a very disadvantageous if implemented with future technologies
low-power interpolator filter between the DSP and the and/or if used for future communication standards.
DAC, but requires also a high-order, power-hungry analog
II. PROPOSED TRANSMITTER ANALOG BASEBAND CHANNEL
reconstruction filter;
* a high DAC conversion frequency, which implies a digital The considerations given in the previous section lead to
filter with a high interpolation factor, but relaxes the a novel TX baseband channel design approach. The solution
required performance of the analog smoothing filter. proposed in this paper is shown in Fig. 2. In this case the
This trade-off is presently optimized with a DAC data-rate input bit-stream at Fc is interpolated to a kFc rate, with k
about 8-to-10 times the signal bandwidth and a 4th-to-6th equal to 30. This operation will require lower and lower power
order analog reconstruction filter. For instance, in the case consumption in scaled technologies, and the circuit portability
of the WLAN standard (whose signal bandwidth is equal to will be immediate, due to its digital nature. On the other hand,
10 MHz), the DAC data-rate is around one hundred MHz [1,2]. this increase of the data-rate/signal-bandwidth ratio would
Looking at future implementations, the above structure require a lower smoothing filter order. In this way the system
will have to face several critical points due to the CMOS is becoming more digital, as the general SoC trend is, since
technology scaling and to the improved standards under def- the analog filter simplification is exchanged with a digital part
inition (802.16 and 802.11n, for instance). The technology complication.
The baseband section architecture of Fig. 2 consists of a
digital interpolator filter from 100 MHz to 600 MHz and
an 8-bit DAC operating at 600 MHz. In this way the use
DSP N
~~U DAC 9 MHzf i _ Active
: 100
8-10bit <t_ 4th6th
filter
@, secio
orde ~~~~section of an analog active reconstruction filter is avoided, result-
digital section baseband analog section ing in a significant power consumption reduction. A single
pole filtering (provided by the DAC passive load), combined
Fig. 1. Traditional baseband analog section for wireless transmitters with the "sinc" attenuation inherently provided by the DAC,
ToRF
suitably suppresses the images around the multiples of a 1.section
conversion frequency as high as 600 MHz. Considering for
example the signal bandwidth of 10 MHz and 14 MHz for
WLAN 802.1 la/g and 802.16 respectively, a DAC output pole
frequency set to 40 MHz gives a total attenuation of the first
signal replica greater than 55 dB. This ensures that all the DAC GND Unit cell
images lie under the emission mask. Moreover the resulting
large oversampling ratio determines an improvement in the Fig. 3. Architecture of the implemented DAC
linearity performance due to the reduced amplitude of the
transitions between successive codes.
The main drawback of this architecture may be the dis- voltage swing of 260 mVpp on each output node does not affect
tributed noise due to the high operating frequency. However the linearity of the DAC thanks to the high output impedance
this is mitigated by three considerations: the low number of of the current cells. This is achieved by the cascoding action
output bits determines a very small interpolator section, the of the switches in the on state.
large OSR reduces the number of cells effectively switching IV. EXPERIMENTAL RESULTS
at 600 MHz, and the technology shielding reduces crosstalk.
Moreover, differently than in the receiver, in the transmitter a The DAC system is realized in a standard 0.13-[tm CMOS
significant signal level is present, reducing the dynamic range technology. The chip microphotograph is shown in Fig. 4. The
requirements. The experimental results demonstrate that this DAC consumes 2.4 mW from a single 1.2-V supply voltage
distributed noise is not critical. and occupies 0.27 mm2. The 100-MHz to 600-MHz interpo-
lator filter has been designed and it can be implemented by
III. DAC ARCHITECTURE means of a 2-kgate block with two 10-stages FIR digital filters
to suppress the baseband images under the emission mask. The
Fig. 3 shows the architecture of the proposed DAC system, interpolator filter area is evaluated in about 0.2 mm2, while
operating from the challenging 1.2 V supply voltage. A fully the estimated power consumption is 2.3 mW. The proposed
thermometric differential current-steering structure guarantees solution achieves an improvement of more than 50% in power
high static and dynamic linearity performance. The low reso- consumption with respect to the typical solutions [3].
lution implies a negligible area of the binary-to-thermometric In the chip of Fig. 4 a dedicated digital part used for
decoder. The unit cell, with a current of 5 [tA, is sized to high-frequency testing has been included in the device, and it
reduce the effect of random mismatch. A relative standard occupies the largest amount of area (24 kgate). The crosstalk
deviation in the current value equal to 0.7% ensures an INL between the digital and the analog part is kept as low as
yield of 99% with a 0.5LSB limit. The overall dimensions of possible using a shielding triple well, separated power supplies
the current sources matrix (0.05 mm2) are such that systematic and multiple bonding wires to reduce parasitic inductances.
errors due to process variations do not affect the device The achieved performance immunity to the crosstalk demon-
linearity. The switches, which deviate the unit current towards strates the robustness of the proposed structure with respect to
the positive or negative output, are driven with low-swing
signals (from 300 mV to 800 mV) to limit the energy of
glitches. Their sizes are kept to minimum values to reduce
the charge injection effects during the switching phase. Two
drivers, placed after the synchronization latches, are used to
generate the desired voltage levels. Each driver is biased with
a current of 100 [A.
The passive load converts the output currents into a voltage
signal and filters the signal replicas. In Fig. 3, a resistance R
equal to 220 Q in parallel with a capacitance C of 17.5 pF
implements the required 40-MHz single pole transfer function.
A common-mode feedback circuit, which acts on the two fixed
PMOS current sources, adjusts the voltage of the common
node A of the two impedances to 600 mV. The single-ended Fig. 4. Chip microphotograph
405
70-
68
66
64
62
> 60
m 58
D 56
54
52
50
48-
0
-
5 10 15
Input signal frequency (MHz)
20
coupling disturbs.
The proposed DAC operates with a conversion frequency of
600 MHz. For a full-scale (FS) single tone with a frequency
Fin up to 20 MHz (which accounts also for the WLAN
802.11n standard under development) Fig. 5 shows that the
third harmonic distortion (HD3) is higher than 59 dB over the
whole input frequency range. In the same Fig. 5, the measured
IMD3 performance, for two -6-dBFS input tones with AF
equal to 1MHz and center frequency up to 20 MHz, is given.
The integrated power spectral density (evaluated with a
static code at the DAC input) results in -57 dBm. Conse-
quently, the FS dynamic range (DR) of the system, obtained
adding the contribution of the 8-bit quantization noise filtered
by the DAC single pole, is equal to about 52.9 dB (8.5 bits).
Fig. 6 shows the output spectrum of a 9-MHz FS tone. The
measured attenuation deriving from the DAC "sinc" filtering
-=20
_|1_ f
0i
--2
0
--3
l11l 0
-
0 _ rX*1|
Inpu
-4
en-5
28C0
ton
1_idB,,
4
u
}5NH
Att
bC 61 ;, t * I .. _ _ igna
60 t 11 = = oP§ls41W 30 kT-I:
____-_*
SliT
Fig. 7. Spectrum of two -6-dB FS tones at 13.6 MHz and 15.4 MHz
and the single pole filtering at the frequency of the first replica performs the baseband digital-to-analog processing with the
corresponds to the one expected from design. Fig. 7 shows the highest efficiency. Moreover it is the one operating with a
output spectrum for an intermodulation test with the two tones supply voltage as low as 1.2 V, while guaranteeing the standard
center frequency set to 14.5 MHz. Fig. 8 shows the output requirements.
spectrum when one component of a WLAN 802.1 la signal
is applied at the DAC input. As expected, the first replica of V. CONCLUSION
3
_lvi kH.
rXasmai m-s
1 /
_
l,-
|
r- _li_
_
9
p:l79IU
|11
the signal around the conversion frequency remains under the A DAC system to be embedded in WLAN 802.11 and
level indicated by the standard emission mask. 802.16 transmitters is presented. The circuit exploits the over-
The performance summary is reported in Table I. A com- sampling ratio to avoid the use of the analog reconstruction
parison with other recent published DACs can be made using filter following the DAC. The DAC is fabricated in a standard
the Figure of Merit reported in [4]: 0.13-[tm CMOS technology and operates with a conversion
frequency of 600 MHz, consuming 2.4 mW from a 1.2 V
Vswing
FOM = .Fin 10SFDR/20 (1) power supply. The measured SFDR is greater than 59 dB for
an input frequency up to 20 MHz, which accounts also for
where Vswing is the output differential peak-to-peak swing next generation wireless protocols, as WLAN 802.11n and
expressed in V, P is the power consumption expressed in mW, 802.16. The IMD3 is higher than 50 dB in the same frequency
Fin is the frequency in kHz at which the SFDR (expressed in range. The dynamic range in a band of 10-MHz is 60.6 dB,
dB) is measured. Fig. 9 shows that, for signal frequencies corresponding to an effective resolution of 9.7 bits. These data
of 10 MHz and 14 MHz, which account for the WLAN correspond to the highest efficiency baseband digital-to-analog
802.1 la/g and 802.16 targets respectively, the proposed block conversion if compared with state-of-the-art DACs.
406
x 106
.
o
=., | WLAN baseband signal * FOM at 1OMHz
6 this work
-1 El FOM at 14MHz
10 --- - --- --1 '- -- ------ -- --- -- --
20 5 * -this work
-;
30-- -:- :--:- :-::N
:-::----m-i
-W -m
-a
\4\WLAN emission mask- Cong. (ISSCC3)
Go
4n
4u [ :::::\ :-::------
a)33-
:E ...
Clara' (ISSCC05)
.. ....
O' Sullivan (JSSCC04)
-50
.O2. * Clara (ISSCC05) 0
-60 Schofield
C:l Deveugele (ISSCC04) * ISSCCO3)
' Doris. (iSiSCCO}5). IS
-70 * Hyde (JSSCC03)
r_ Huang (ISSCC04)
Chen (ESSCIRC04)
-80 L II
10 100 700 15 18 5.5 33
Frequency (MHz) knalog supply voltage [V]
Fig. 8. Spectrum of a component of a WLAN 802.1la signal Fig. 9. Figure of Merit vs analog supply voltage for recently published DACs
TABLE I TABLE II
PERFORMANCE SUMMARY. PARAMETERS USED FOR THE EVALUATION OF FOM FOR RECENTLY PUBLISHED
DACS AND FOR THE IMPLEMENTED ONE.
Parameter ] Value
Technology CMOS 0.13 tm P Vswing SFDR [dB] SFDR [dB]
Supply voltage 1.2 V [mW] I[V] @14MHz @10MHz
Core area 0.27 mm2 This work 2.4 0.52 65.5 67
Maximum conversion rate 600 MS/s Cong [5] 16.7 1 72 76
Power consumption 2.4 mW Clara [6] 45 1.536 76 76
Differential FS output swing 520 mVpp O'Sullivan [7] 82 2 77 80
INL/DNL <0.25LSB Doris [8] 160 1.5 80 80
SFDR @ FS, Fin = 12 MHz 68 dB Schofield [9] 400 1 90 95
F, = 600 MHz Deveugele [10] 22 0.5 70 73
IMD3 @ -6dBFS, F,enter = 12 MHz, 57 dB Hyde [11] 53 1 72 72
AF 1MHz, F, = 600 MHz Huang [12] 97 0.8 72 73
DR @ FS 52.9 dB (8.5 bits) Chen [13] 103 5 45 49
DR @ FS, 0-10 MHz 60.6 dB (9.7 bits)
[7] K. O'Sullivan et al., "A 12b 32OMsample/s Current Steering CMOS D/A
Converter in 0.44 mm2," IEEE Journal of Solid-State Circuits, pp. 1064-
ACKNOWLEDGMENT 10722, July 2004.
[8] K. Doris et al., "A 12b 500MS/s DAC with >70dB SFDR up to 120MHz
This research has been partially supported by the Italian in 0.18 Ftm CMOS," ISSCC Dig. of Tech. Papers, pp. 116-117, February
National Program FIRB, Contract n° RBNEOlF582. 2005.
[9] W. Schofield et al., "A 16b 400MS/s DAC with <-8OdBc IMD to 300MHz
and <-l6OdBm/Hz Noise Power Spectral Density," ISSCC Dig. of Tech.
REFERENCES Papers, pp. 126-127, February 2003.
[10] J. Deveugele and M. Steyaert, "A 10b 250MS/s Binary-Weighted
[1] S. Mehta et al., "An 802.11g WLAN SoC," ISSCC Dig. of Tech. Papers, Current-Steering DAC," ISSCC Dig. of Tech. Papers, pp. 362-363, Febru-
pp. 94-95, February 2005. ary 2004.
[2] H. Darabi et al., "A Fully Integrated Soc for 802.1 lb in 0.18 rm CMOS," [11] J. Hyde et al., "A 300-MS/s 14-bit Digital-to-Analog Converter in Logic
ISSCC Dig. of Tech. Papers, pp. 96-97, February 2005. CMOS," IEEE Journal of Solid-State Circuits, pp. 734-740, vol. 38, no.
[3] N. Ghittori et al., "A Low-Power, Low-Voltage (1lmW/8.4mW, 1.2V) 5, May 2003.
DAC+Filter for Multistandard (WLAN/UMTS) Transmitters," VLSI Dig. [12] Q. Huang et al., "A 200MS/s 14b 97mW DAC in 0.18km CMOS,"
of Tech. Papers, pp. 334-337, June 2005. ISSCC Dig. of Tech. Papers, pp. 364-365, February 2004.
[4] D. Giotta et al., "Low-Power, 14-bit Current Steering DAC, for [13] T. Chen et al., "A 14-Bit 130-MHz CMOS Current-Steering DAC with
ADSL2+/CO Applications in 0.13 Ftm CMOS," Proc. of ESSCIRC, pp. Adjustable INL," Proc. of ESSCIRC, pp. 167-170, September 2004.
163-166, September 2004.
[5] Y Cong and R. L. Geiger, "A 1.5V 14b IOOMS/s Self-Calibrated DAC,"
ISSCC Dig. of Tech. Papers, pp. 128-129, February 2003.
[6] M. Clara et al., "A 350MHz low-OSR AY Current-Steering DAC with
Active Termination in 0.13 Ftm CMOS," ISSCC Dig. of Tech. Papers, pp.
118-119, February 2005.
407
A 1.8V 20mW 1mm² 14b 100MSample/s CMOS DAC
Mika Tiilikainen
Nokia Mobile Phones, Helsinki, Finland
mika.p.tiilikainen@nokia.com
0 CB mean(CA,CB) CA code
Figure 4. Principle of the two-phase calibration. The CALDACS
dark lines represent the actual currents seen by the
LSB MID MSB
current comparator.
SWITCHES
4. Experimental results CALMIRRORS
LATCHES
A 14-bit DAC was implemented in a digital CMOS
process and the full linearity was achieved with the
Figure 5. Layout of the DAC.
calibration. The layout of the DAC is presented in Fig. 5.
The micrograph of the chip would be useless, because
the deep-submicron process is equalized by covering the
4
whole chip with dummy metal plates. The 6-bit
INL [LSB]
2
calibration DACs are oversized and could be reduced
0
significantly resulting in a smaller die area. As a result,
−2
the calibration increases the area of the DAC
−4
approximately 50%. It is not significant, because the area
1 4096 8192 12288 16384
of the DAC itself is relatively small for many code
applications. The floorplan is basically the same as used 2
e.g. in [2], where the latches are isolated from the analog 0
DNL [LSB]
parts. −2
In this study, the main emphasis was on the static
−4
linearity problems of a binary weighted DAC. As shown
in Fig. 6, the effective number of bits (ENOB) would −6
1 4096 8192 12288 16384
have been approximately ten without the calibration. code
After trimming, the nonlinearities were reduced to less Figure 6. Static linearity without the calibration.
than half a LSB, as can be seen in Fig. 7. The variations
5. Conclusion
1
A 14-bit binary weighted current steering DAC was
0.5 implemented in a digital process without using any
INL [LSB]
summarized in Table 1.
0 The use of a deep-submicron technology does not
−0.5 reduce the analog area of the DAC significantly, as the
size of the analog components cannot be shrunk
−1
1 4096 8192 12288 16384 appreciably because of the matching requirements. Thus
code
the results are valid also for technologies with larger
Figure 7. Static linearity with the calibration. minimum gate lengths.
p,.
D'Acap T"
I
Off Chip On Chip
LSB * * - MSB
InitializingCapacitor (a)
Other half of
differentialcircult
mz
ringing.
(a)
I
I
I I
I T'- Pipeline Register with DAC cell prechargesla unknown voltage
(b) single-phase clack
Figure 1. (a) Three adjacent unit cells at the LSB end of DDFS Output Word
' \ \
'\\ I!, \l l
(1
I
$2
I I - f + + +
'\\ l
$1
Figure 2. Layout of switch Figure 5. (a) Illustrating the problem caused by clock
FET to lower the stray skew when DAC samples DDFS output. (b) Skew tolerant
capacitance across the OFF clock scheme in pipeline register.
switch.
T
Previous design Faster design
Vcmfb
vref'pi?
*
CMFB
70 70 70
65 65
60
m 60
60 m
-U
m
2 * 55
a
2
=-55
cn
e
m 50
50
50 45
45
40
40 0 30 60 90 120 150
0 25 50 75 100 125
40 SynthesizedFreq, MHz Synthesized Freq, MHz
0.0 0.1 0.2 0.3 0.4 0.5
fsidfclk Figure 10. Comparison of SFDR vs. synthesized sinewave
frequency
(4
(b)
Figure 8. (a) Measured SFDR vs. Normalized synthesized
Clock Rate, MS/s
sinewave frequency (fsin) at various sample rates (fclk).
(b) Non-exponential settling transient arising from Figure 11. Comparison of SFDR (worst-case over Nyquist
nonlinear I-V characteristic of OTA. band) versus sample rate.
20.2 A 10b 250MS/s Binary-Weighted the area of the negative deviations. Reducing the glitch energy to
zero does not eliminate all spurious responses in the Nyquist
Current-Steering DAC
band, but pushes most of the energy out of the band. Fast tran-
sitions push the spurious outputs to even higher frequencies,
Jurgen Deveugele, Michiel Steyaert further reducing the impact in the Nyquist band. In addition,
ESAT-MICAS, Katholieke Universiteit Leuven, Belgium fast switching can decrease the transition energy. The charge
injected by the capacitive coupling of the digital signals to the
The presented 10-bit binary weighted current-steering DAC has output is independent of the switching speed, but the time of
unbalanced switching is reduced. Therefore we designed this
over 60dB SFDR at 250MS/s for signals from dc to Nyquist. The
DAC for minimal glitch energy and as fast as possible switching.
chip draws 4mW from a dual 1.5/1.8V supply plus load currents.
The active area is less than 0.35mm2 in a standard 1P-5M 0.18µm
The latch composed of transistors M1-M6 and inverters I1-I2 can
1.8V CMOS process. Both INL and DNL are below 0.1 LSB.
be seen in Fig. 20.2.2. Transistors M2-M5 and inverters I1-I2
form a David-Goliath latch. It is fast and gives good control over
The degree of segmentation of a current-steering DAC has a
the glitch energy in all process corners. The latch has one pitfall.
major influence on both the structure of the converter and on its
During switching, nodes 1 and 2 are shorted, as are nodes 3 and
performance. Segmented rather then binary weighted architec-
4. A large glitch occurs at these nodes if the logic values differed
tures are often used for improved DNL, glitch, and SFDR per-
before switching. A buffer designed to drive the capacitances on
formance [1]. This however comes at a price. An increase in the
nodes 2 and 3 has too small a capacitance, and extra capacitance
number of unary decoded bits has an exponential influence on
is added at nodes 1 and 4. This increases the power consumption
the area and increases the complexity of the decoder. This
a little, but significantly improves the SFDR. It is essential that
increases the power consumption and the required design time.
all buffers are scaled identically. This ensures that the glitch at
The additional benefit of decoding one extra unary bit gets small
nodes 2 and 3 is signal independent. This is also important for
if the number of decoded bits is high. Even worse, as the digital
segmented structures in which each decoder stage has a differ-
noise increases and time skews get bigger in large decoders, the
ent structure and capacitance, leading to code dependency.
dynamic performance may drop. It is clear that a good segmen- Adding a deglitch stage can have the same effect, but at an extra
tation choice is important. Yet this choice is often based on expe- cost, and the ratio of the capacitances must still be optimized
rience rather than on founded decisions. We have carefully stud- theretoo. Note that many other latches have comparable critical
ied this trade-off and pushed our design to a binary weighted nodes.
structure. The DAC has a highly regular structure, good dynam-
ic performance and low power consumption. Transistors M7-M8 set the voltages on nodes 5 and 6. The node
above the on switch is set by bias voltage Bias3 and the current.
In a segmented structure the DNL is often big at the boundaries The influence of the output voltage is reduced by the gain of the
of the unary and the binary decoded parts. In [2], the linearity cascode device. The node above the off switch is charged expo-
problem is solved by using calibration. It can be solved also if all nentially until the transistor on top of the switch enters sub-
current sources are composites of the same unit cell. This threshold. This charging is complete at the end of one clock peri-
removes the systematic error typically caused by using different od at 250MS/s. The cascode thus shields the switches from the
structures for the MSB and the LSB parts. Therefore the source output voltage variation, and simplifies the design of the driver-
array consists of 1023 identical sources plus the current mirror switch-latch combination for low glitch energy.
and dummies. The DNL specifications are met if one designs for
the correct INL yield (99.7% yield for INL < 0.5 LSB) [3]. The All measurements are done at 250MS/s with 10mA load current
measured INL and DNL plots can be seen in Fig. 20.2.1. unless explicitly stated. Figure 20.2.3 shows the spectrum for a
122.5MHz signal for which the SFDR is 62.3dB. Figure 20.2.4
For good dynamic performance, it is essential to use equal struc- summarizes the spectral performance. The SFDR is over 60dB
tures whenever possible, and switching must be identical for all for all measured frequencies. The DAC was also characterized
big current sources. The chip presented here uses a pseudo-seg- with a load current of 5mA. For this condition, the SFDR
mented structure with 6 pseudo-unary bits and 4 binary bits. dropped by less than 3dB over the Nyquist band. Figure 20.2.5
Therefore the structure is identical to that of a 10-bit 6+4 seg- shows a two tone test with sine waves at 100 and 102.5MHz. The
mented DAC. The inputs are, however, not decoded but buffered SFDR in a 17.5MHz wide band is 67.8dB. Figure 20.2.6 summa-
locally. The MSB consists of parallel connected pseudo-unary rizes the specifications. Of the power, 2mW out of 4mW is used
cells distributed over the switching core. by the clock buffering and buffering at the bondpaths. Figure
20.2.7 shows a die photograph. The chip is bondpath limited, and
When it comes to dynamic performance, there are two main dif- the active area is less than 0.35mm2. Due to the highly regular
ferences between a pseudo-segmented (binary) DAC and a seg- binary structure, the drivers, switches and latches consume very
mented converter. The first is the lack of a decoder. Since the little area.
buffering requires little power, less digital noise is injected into
the output. The second difference is that in a binary DAC a lot of Acknowledgments:
undesirable switching of sources occurs. For example, at major We thank Robert Taft and Chris Menkus from National Semiconductor.
This work has been sponsored and manufactured by National
code transitions in a 10b DAC y = 0 to 511 extra switches are Semiconductor.
switched on, while a similar number of switches are switched off.
However, if the sum of the differential output currents IL an IR References:
remains constant during switching, then this switching is not [1] C. H. Lin and K. Bult, “A 10-b, 500-MSamples/s CMOS DAC in
seen at the output, and the performance of the binary DAC is 0.6mm2,” IEEE J. Solid-State Circuits, vol. SC-33, no. 12, pp. 1948-1958,
then equal to that of the unary DAC. The deviations from the Dec 1998.
[2] M. P. Tiilikainen, “A 14-bit 1.8-V 20 mW 1- mm2 CMOS DAC,” IEEE J.
ideal must be minimized.
Solid-State Circuits, vol. SC-36, no. 7, pp. 1144-1148, Jul 2001.
[3] M. Borremans, A. V. D. Bosch, M. Steyaert, and W. Sansen, “A Low
One way to state this problem is to say that the glitch energy has Power 10-bit CMOS D/A Converter for High Speed Applications,” IEEE
to be minimized. Then the area of the positive deviations equals CICC Digest, pp. 265-268, May 2000.
Figure 20.2.1: Measured INL and DNL plot. Figure 20.2.2: Schematics of a single buffer-latch-switch cell.
Figure 20.2.4: SFDR plot for measured frequencies. The sample rate is
Figure 20.2.3: Measured SFDR plot for a 122.5 Mhz signal at 250MS/s. 250MS/s.
Resolution 10 bit
Decoding binary
Nyquist update rate 250 MS/s
INL < 0.1 LSB
DNL < 0.1 LSB
Differential signal, Iload = 10 mA 500 mVpp
Active area < 0.35 mm2
SFDR up to Nyquist, Iload = 10 mA > 60 dB
SFDR up to Nyquist, Iload = 5 mA > 57.8 dB
Glitch energy, Iload = 10 mA 2.64 pV.s
Power@Nyquist 4 mW + Iload x 1.8 V
Voltage supply 1.5 V and 1.8 V
Technology 0.18 Pm CMOS 1P 5M
Figure 20.2.5: Two tone measurement. Tones at 100 and 102.5MHz. The
SFDR in the shown band is 66.76dB. Figure 20.2.6: Performance specifications.
Abstract—This paper studies the impact of segmentation on cur- way to boost dynamic performance [3] may ignore better ways
rent-steering digital-to-analog converters (DACs). Segmentation of dealing with dynamic nonlinearities.
may be used to improve the dynamic behavior of the converter Reducing the segmentation degree may thus be advantageous
but comes at a cost. A method for reducing the segmentation
degree is given. The presented chip, a 10-bit binary-weighted for the overall performance of the converter. Although some
current-steering DAC, has 60 dB SFDR at 250 MS/s from DC papers discuss the importance of segmentation [3], it is still
to Nyquist. At 62.5 MHz signal frequency and 250 MS/s, we poorly understood. Rather than assuming that extra segmenta-
operated the device in 9-bit unary, 1-bit binary-weighted mode. tion improves the performance, we tried to separate the effects
The obtained 60 dB SFDR in this measurement demonstrates caused by segmentation from the effects caused by other mech-
that the binary nature of the converter did not limit the SFDR.
The chip draws 4 mW from a dual 1.5 V/1.8 V supply plus load anisms. Soon we realized that very few mechanisms for gen-
currents. The active area is less than 0.35 mm2 in a standard erating spurious signals are strongly related to the segmenta-
1P-5M 0.18- m 1.8-V CMOS process. Both INL and DNL are tion. Addressing these mechanisms could thus yield a binary-
below 0.1 LSB. weighted DAC with high SFDR at high signal and clock fre-
Index Terms—Binary weighted, digital-to-analog converters, quencies. So, we designed a binary-weighted converter rather
low power. than a segmented one. We see that the binary-weighted con-
verter is still less tolerant to poor design than the segmented
I. INTRODUCTION one, so we chose, for simplicity, to use generous margins on all
parameters.
Fig. 2. One quarter of the switching scheme used. Current sources with
numbers 1 to 15 are used for the pseudo-segments. Current sources with
number 0 are used for the LSB part. The scheme is tolerant to gradient errors
Fig. 1. Yield as a function of the standard deviation on the current sources and reduces the impact of the edges of the current-source array.
for both unary and binary-weighted converters.
Fig. 3. Scheme for a segmented converter. The converter has a decoding stage,
a switching core with the switches, cascodes and latches, and a current-source
array.
Fig. 5. The glitch energy equals to the energy difference between the ideal
transition and the real transition. Most of the energy is due to the slower than
ideal transition.
Fig. 11. Simulation of the influence of the output voltage on the SFDR,
unary-decoded converter is more susceptible to this effect. The differential-ended. (upper plot) Unary-weighted converter. (lower plot)
effect of the different up and down transitions has even-order Binary-weighted converter.
distortions only. Therefore, it is not necessary to simulate the
differential case. is modeled to be dependent on the mean value of the output be-
fore and after switching. We make this assumption because we
B. Influence of the Output Voltage on the SFDR believe the rise time is dependent on both the output voltage
A high output voltage is preferred as it increases the signal at the beginning of the transition and on the output voltage at
power. The output voltage, however, is likely to influence the the end of the transition. If the code is decreasing, we make the
transitions from one code to another. Transistors M7–M10 in fall time dependent on the output voltage of the complementary
Fig. 6 are operated in the saturation region. They act as cas- output in order to keep the fully differential nature of the struc-
codes between nodes N7, N8, and N9. Due to the finite gain of ture.
the transistors some part of the output voltage is seen on node Fig. 10 shows the influence of the output voltage on the SFDR
N5–N6 and N9. This influences the timing of the switching. In for the single-ended case. The spurious signals caused by the
our model, we define the transition waveforms to be exponen- output voltage are even and odd order. The results for the dif-
tial functions that have time constants that are dependent on the ferential case are shown in Fig. 11. Here again, we see that the
output voltage. If the input code is increasing, then the rise time unary-decoded converter outperforms the binary-decoded at low
DEVEUGELE AND STEYAERT: A 10-bit 250-MS/s BINARY-WEIGHTED CURRENT-STEERING DAC 325
V. IMPLEMENTATION
Our analysis (and the analysis in [7]) both indicate that there
is no clear reason why a unary (or segmented) converter outper-
forms a binary-weighted converter by an order of magnitude in
published designs. We also showed that transition times must
be as fast as possible. We therefore opt to make a pseudo-seg-
mented converter with as fast as possible transition times. A
pseudo-segmented converter is identical to a segmented one, ex-
cept that the bits at the input of the segments are decoded binary
instead of unary. Fig. 13. Example simulated waveform when the clock turns high.
Fig. 15. Mirrored layout of parallel inverter instances reduces the area requirements and the parasitic capacitances.
Fig. 18. SFDR plot for measured frequencies. The sample rate is 250 MS/s.
Fig. 20. Measured spectrum for a 0.01 2 full-scale signal around the midcode.
Fig. 19. Two-tone measurement. Tones are at 100 and 102.5 MHz. The SFDR
in the shown band is 66.76 dB.
TABLE I
SPECIFICATIONS
TABLE II
COMPARISON WITH OTHER BINARY-WEIGHTED DACS
Fig. 22. Chip photograph. The current-source array and the drivers, switches,
and latches are indicated. The chip is bond-path limited. Most of the area is filled
with dummies.
fast and compact current-steering converters that scale well in [11] M. P. Tiilikainen, “A 14-bit 1.8-V 20-mW 1-mm CMOS DAC,” IEEE
deep-submicron technologies. J. Solid-State Circuits, vol. 36, no. 7, pp. 1144–1147, Jul. 2001.
ACKNOWLEDGMENT
The authors thank R. Taft and C. Menkus for their support at Jurgen Deveugele (S’01–M’05) received the
Master’s degree in electronic engineering and the
tape-out. Ph.D. degree in electronics from the Katholieke Uni-
versiteit Leuven (K.U.Leuven), Heverlee, Belgium,
in 1999 and 2005, respectively.
REFERENCES His main research interests are in low-power
[1] A. V. den Bosch, M. Borremans, M. Steyaert, and W. Sansen, “A 10-bit and high-speed current-steering digital-to-analog
1-G sample/s Nyquist current-steering CMOS D/A converter,” IEEE J. converters.
Solid-State Circuits, vol. 36, no. 3, pp. 315–324, Mar. 2001.
[2] M. Borremans, A. V. den Bosch, M. Steyaert, and W. Sansen, “A low
power, 10-bit CMOS D/A converter for high speed applications,” in
Proc. IEEE Custom Integrated Circuits Conf. (CICC), May 2001, pp.
157–160.
[3] C.-H. Lin and K. Bull, “A 10 b, 500 M sample/s CMOS DAC in 0.6
mm ,” IEEE J. Solid-State Circuits, vol. 33, no. 12, pp. 1948–1958, Dec. Michiel S. J. Steyaert (S’85–A’89–SM’92–F’04)
1998. received the Master’s degree in electrical-mechanical
[4] A. Bugeja, B. Song, P. Rakers, and S. Gillig, “A 14-b, 100-MS/s CMOS engineering and the Ph.D. degree in electronics from
DAC designed for spectral performance,” IEEE J. Solid-State Circuits, the Katholieke Universiteit Leuven (K.U.Leuven),
vol. 34, no. 12, pp. 1719–1732, Dec. 1999. Heverlee, Belgium, in 1983 and 1987, respectively.
[5] G. V. der Plas, J. Vandenbussche, W. Sansen, M. Steyaert, and G. Gielen, From 1983 to 1986, he obtained an IWNOL fel-
“A 14-bit intrinsic accuracy Q random walk CMOS DAC,” IEEE J. lowship (Belgian National Fundation for Industrial
Solid-State Circuits, vol. 34, no. 12, pp. 1708–1718, Dec. 1999. Research) which allowed him to work as a Research
[6] J. Deveugele, G. V. der Plas, M. Steyaert, G. Gielen, and W. Sansen, Assistant at the Laboratory ESAT at K.U.Leuven.
“A gradient-error and edge-effect tolerant switching scheme for a high- In 1987, he was responsible for several industrial
accuracy DAC,” IEEE Trans. Circuits Syst. I: Fund. Theory Appl., vol. projects in the field of analog micropower circuits
51, no. 1, pp. 191–195, Jan. 2004. at the Laboratory ESAT as an IWONL Project Researcher. In 1988, he was a
[7] K. Doris, C. Lin, and A. van Roermund, “Synchronization distortion and Visiting Assistant Professor at the University of California, Los Angeles. In
spatio-temporal switching in current steering D/A converters,” in Proc. 1989, he was appointed by the National Fund of Scientific Research (Belgium)
IEEE Benelux Workshop on Circuits, Systems and Signal Processing as a Research Associate, in 1992, as a Senior Research Associate, and in 1996,
(ProRISC), The Netherlands, Nov. 2000, pp. 259–266. as a Research Director at the Laboratory ESAT, K.U.Leuven. Between 1989 and
[8] B. Schafferer and R. Adams, “A 3 V CMOS 400 mW 14 b 1.4 GS/s 1996, he was also a part-time Associate Professor. He is now a Full Professor
DAC for multi-carrier applications,” in IEEE Int. Solid-State Circuits at the K.U.Leuven. His current research interests are in high-performance and
Conf. (ISSCC) Dig. Tech. Papers, vol. 47, Feb. 2004, pp. 360–361. high-frequency analog integrated circuits for telecommunication systems and
[9] Q. Huang, P. A. Francese, C. Martelli, and J. Nielsen, “A 200 Ms/s 14 analog signal processing.
b 97 mW DAC in 0.18 m CMOS,” in IEEE Int. Solid-State Circuits Prof. Steyaert received the 1990 and 2001 European Solid-State Circuits Con-
Conf. (ISSCC) Dig. Tech. Papers, Feb. 2004, pp. 364–365. ference Best Paper Award. He received the 1991 and the 2000 NFWO Alcatel-
[10] K. O’Sullivan, C. Gorman, M. Hennessy, and V. Callaghan, “A 12-bit Bell-Telephone award for innovative work in integrated circuits for telecommu-
320-M sample/s current-steering CMOS D/A converter in 0.44 mm ,” nications. He received the 1995 and 1997 IEEE ISSCC Evening Session Award
IEEE J. Solid-State Circuits, vol. 39, no. 7, pp. 1064–1072, Jul. 2004. and the 1999 IEEE Circuit and Systems Society Guillemin–Cauer Award.
A 1-V 2.5-mW Transient-Improved Current-Steering DAC
using Charge-Removal-Replacement Technique
Ka-Hou Ao leong, Seng-Pan U 1 and R.P. Martins 2
Analog and Mixed-Signal VLSI Laboratory, FST, University of Macau, Macao, China
1-Also with Chipidea Microelectronics (Macao) Ltd., 2-On leave from Instituto Superior Tecnico (IST)/UTL, Lisbon, Portugal
Abstract - A charge-removal-replacement (CRR) technique is
proposed to realize a low-voltage low-power current-steering
digital-to-analog converter (DAC) for minimizing transition time
in current switching. Implemented in a 0.18-pm CMOS process,
the 10-bit 120-MS/s DAC focused on WLAN applications
consumes only 2.5 mW from a single 1-V supply. The simulated
monotonic performances achieve a mean value of IINLI < 0.14
LSB and IDNLI < 0.18 LSB, respectively. The averaged SFDR
obtains 67.23/62.45 dB with/without the CRR technique.
I. INTRODUCTION
The demand for high performance digital-to-analog converters
(DACs) has obviously increased for the fast development of (a) (b)
Fig. 1 (a) Cascoded-current-source and (b) cascoded-switch topology.
advanced telecommunication systems, e.g., wireless local area
networks (WLANs). In view of that, most of data converters are
desired to be fully integrated into a CMOS System-On-a-Chip
(SoC) wireless transceiver with including RF transceiver, digital
physical layer (PHY), corresponding media access controller Vm+rnt CRR , OFF
(MAC) and all other analog baseband building blocks [1]. To
avoid any supply-voltage constraints for the technology scaling,
low-voltage building blocks with low-power consumption is Ml' - ONH
highly required. vssvss ~~~~~~~~~~~~Time
In practice, current-steering DAC is preferred for wireless
communications due to its high sample rate, capability of driving
resistive loads and relatively low-power consumptions. The
current cells with cascoded topologies as shown in Fig. 1 are
satisfactory for static and dynamic output-impedance requirements
[2]-[3]. Alternatively, the cascoded-switch topology [4] presents as
additional digital signal feedthrough isolation and output influence nVwell nAwell A VSSTime
reduction through the insertion of two cascoded transistors (M,a,+ (a) (b)
and M, -). However, asymmetry in rising-falling transitions and Fig. 2 (a) Operational principles of CRR technique and (b) the
long settling time will degrade the dynamic performance directly, corresponding control signals.
especially applied in high speed and high resolution DAC which
unit current is small. This paper proposes a charge-removal- cascoded switch (Me.,lVLa) enters in the subthreshold region. As
replacement (CRR) technique that fixed those transient problems. a result, an amount of current leaks into the path even if the
For demonstrated purposes, a 1-V 10-bit 120-MS/s DAC with less corresponding switch is in the off stage. Such current is given by
than 1.3-jtA unit current was designed and simulated, the current the subthreshold expression of the drain current (ID),
switching transient is highly reduced and the dynamic
performances of the DAC are dramatically enhanced. ID =Io exp V><(1)
After this introduction, the proposed CRR technique will be where ;>I is a nonideality factor and VT=- kT q. Moreover, with
presented in section II. In section III, it will be applied to a low- parasitic capacitance present in the source terminal of cascoded-
voltage current-steering DAC with a detailed description of each transistor Mc,. Mc,. , the transition will be further extended.
building block. In section IV, simulation results will be presented.
Finally, the conclusions will be summarized in section V. Operational Principles The CRR technique is depicted in
Fig. 2(a), where a capacitor (CcR) driven by an inverter is inserted
II. CHARGE-REMOVAL-REPLACEMENT TECHNIQUE between the gate and drain terminals of the switched transistor.
The main disadvantages of the conventional cascoded-switch When the desired switch is turning off, the inverter generates an
current source (CSCS) topology are the long settling time and inverted control signal to draw out the positive ((IqD)
charge from
asymmetry settling time (falling settling time is slower than that in the inversion layer and the ubiquitous parasitic capacitance at node
rising), which distorting output current of the DAC. The tarry Xb such that any lengthy discharge procedure can be eliminated.
falling settling time is caused by the node voltage, which below To avoid current sources from being turned off simultaneously,
the off switch (MA,,1M,), is discharged exponentially until the digital control signals with lower crossing point apply to the switch
183
1-4244-0387-1/06/$20.00 (@2006 IEEE
1
LU)
0
a)
(a)
1
5
-o
< settle ,,wl CRR
CRR
a *
'a 1
LU, n0.5
D. o -1. \ ,~~~~~~~~,
settle
- 5 10 15 20 25 30 35 40 45
I
50
0 20M 40M 60M 80M 1 OOM 120M
Input Code Rate of Unit Current Cell (Hz)
Time (ns)
Fig. 4 Input code rate vs. instantaneous node voltage vx settling level, and
(b) rising/falling delay of differential output current.
05 a-- settle C
5n
0.4- ..R
not settle |_ .................
(L RR
0)
.................
|0 fl j=j ~~~~~~~~~~.
...
0.3 settle
tt......
0.2._
.~I d_
.......
0 5 10 15 20 25 30 35 40 45 50
o
45.99%
[1
. . . . . . . . . . . . . . . . .~ . . . . . . . .
Time (ns) ......... V rio _C8
tion ..
(c)
Fig. 3 (a) Single-ended output current, (b) differential-ended output .. C ........ ........... ............................
[different] I[average]
below: (t: Calculated by in the range of input code rate.)
charge, given that the node voltage vx is faster settling to its resulting in a non-constant rising delay for switching on the current
steady-state value while the current path is turning on. On the other path. As shown in Fig. 4, the CRR technique reduces the variation
hand, when the current path is desirable turned off, the CRR of vx from the steady-state voltage for different switching rates,
technique removes the residual charge and prompts the node and it also leads the rising/falling delays of the differential output
voltage vx to steady state (i.e., vx is leaded to reduce). Therefore, currents to be more independent of the input code rate. In fact, the
the cascoded switches no longer suffers from the critical difference between rising/falling delays is dependent on the
exponential discharge problem, establishing balanced falling and capacitance of CCRR. With capacitor variation, the performance is
rising speeds. As illustrated in Fig. 3(a), the falling transition of a slightly degraded as shown mn Fig. 5. However, the tmprovement is
single-ended output current is enhanced in terms of speed to still significant when comparing with the performance of
achieve more balanced rising-falling characteristic. Even though a conventional structures. Furthermore, pushing the cascoded
differential-ended output signal [Fig. 3(b)] can suppress even- switches to the deep subthreshold region can eliminate the induced
harmonic distortion in the conventional CSCS structures, the current, such that the accuracy of the D/A conversion will not be
transient nonlinear problem still exists, which can be solved by the degraded.
and increase the susceptibility to asynchronous glitches. The and the output impedance (Z0ut) has two poles and two zeros as:
APCCAS 2006
184
0.1k 10k IM 1O0M 1 OG
Frequency (Hz)
Fig. 6 Output impedance with and without CRR.
30 30 2 0.25
02
20 20
0.15
0.1
10 10
0.05
a -80 a 80 L66
bn nCD
-100 0
10
-120 -120
6
Output..Signa Fre. ....
-1 60' '
0 1OM 20M 30M 40M 50M 60M 0 1OM 20M 30M 40M 50M 60M :::M 1M
Frequency (Hz) Frequency (Hz)
(a) (b) (c)
Fig. 9 Simulated output spectrum at 5 MIHIz, 120 MS/s (a) without and (b) with CRR technique. (c) SFDR vs. output signal frequency.
the output spectrums are presented in Fig. 9 (a) and (b), chip area overheads. In average, 4.78-dB improvement in SFDR is
respectively. With the new technique, the averaged SFDR is achieved within the Nyquist band.
improved by 4.78 dB within the Nyquist band [Fig. 9(c)]. A
summary ofthe simulation performances is listed in Table 1. ACKNOWLEDGMENT
This work was financially supported by the University of
V. CONCLUSION Macau under the research grant with Ref No RG069/02-03S/
This paper presented a charge-removal-replacement (CRR) MR/FST.
technique for improving the transient performances of current- REFERENCES
steering DACs. Adopted in a 1-V 10-bit 120-MS/s fully binary-
weighted current-steering DAC, the technique removes (replaces) [1] S. Mehta, et al., "An 802.11g WLAN SoC," in IEEE ISSCC, Digest of
the node charge when the desired current path is turned off (on). Technuical Papers, pp. 94-96, Feb. 2005.
[2] B. Razavi, Principles of Data Conversion System Design. IEEE Press,
The discharging (charging) time is reduced leading to a higher 1995.
speed in D/A conversion. Compared with the conventional [3] A. V. D. Bosch, et al., "SFDR-Bandwidth Limitations for High Speed
cascoded-switch current source (CSCS) topology, the switching High Resolution Current Steering CMOS D/A Converters," in Proc. of
transition, asymmetry rising-falling settling and synchronization ICECS, VOL. 3, pp. 1193-1196, Sept. 1999.
problems are minimized with negligible power (< 0.15 mW) and [4] H. Takakura, et al., "A 10 bit 80MHz Glitchless CMOS D/A Converter,"
in Proc. of CICC, pp. 26.5.1-26.5.4, May 1991.
Table 1 Summary of the DAC performance. [5] K. -H. Ao leong, et al., "A Frequency Up-Conversion and Two-Step
Channel Selection Embedded CMOS D/A Interface," in Proc. ofISCAS,
Specification Unit Value pp. 392-395, May 2005.
Technology 0.1 8-,tm CMOS [6] N. Ghittori, et al., "A Low-Power, Low-Voltage (1 1mW/8.4mW, 1.2V)
Number ofbits 10 DAC+Filter for Multistandard (WLAN/UMTS) Transmitters," in Proc.
Sample rate MS/s 120 of VLSI Circuits, pp. 334-337, Jun. 2005.
Supply voltage (Analog & Digital) V 1 [7] J. Deveugele, et al., "A 10b 250MS/s Binary-Weighted Current-Steering
Differential output signal V 0.5 DAC," in IEEE ISSCC, Digest of Technical Papers, pp. 362-532, Feb.
Load (RL) Q 200 2004.
JINLI [Mean, Sta Div] LSB [0.134, 0.049] [8] M. J. M. Pelgrom, et al., "Matching properties of MOS transistor", IEEE
IDNLI [Mean, StaDiv] T LSB [0.178,0.081] JSSC, VOL. 24, NO. 5, pp. 1433-1439, Oct. 1989.
SFDR @ Desired signal band j dB > 63 [9] A. V. D. Bosch, et al., "A 10-bit 1-GSample/s Nyquist Current-Steering
Glitch energy pV-s 0.4 CMOS D/A Converter," IEEE JSSC, VOL. 36, NO.3, pp. 315-324, Mar.
Power consumption mW l_ 2.51 l 2001.
t: Power Breakdown: 1.3 mW (Analog), 1.2 mW (Digital).
M i s m a t c h b e t w e e n
a M S B s o u r c e a n d
6 7
T h e r m o m e t e r 6 7 C u r r e n t t h e L S B a r r a y
1 3 C u r r e n t s w i t c h e s
d e c o d e r s w i t c h e s
C o m p a r a t o r
6 7
3 O f f s e t
1 0
1 3 c a l i b r a t i o n
M S B s o u r c e s L S B s o u r c e s
s o u r c e s
2 0
1 6
D i g i t a l I n p u t
C u r r e n t
i n p u t C o n t r o l
s c a l e r
I n p u t C o d e
1 1
S c a l e
C a l i b r a t i o n
I B I A S
c o n t r o l
u n i t
Fig. 2. Calibration principle.
C l o c k
C l o c k +
g e n e r a t o r
C l o c k -
means that the more accurate the MSB array is, the less
c
C
o
l o
n t r
c k
o l
1 2
,
O u t p u t C u r r e n t Res1 Res2 Res3 Sample Res4
Res1
Res2
4 L S B Res3
Sample
Res4
Comp
C A L - v a l u e
M e a s u r e d
CAL-value REF-value
at input at input
o f f s e t
3.4. Comparator
1 / 8 L S B
,
O u t p u t C u r r e n t
INL uncalibrated INL calibrated
0.8
5
4 0.6
3 0.4
INL [LSB]
INL [LSB]
2
0.2
1
R e f e r e n c e 1 / 8 L S B
0 0
v a l u e Ŧ1
Ŧ0.2
Ŧ2
Ŧ3 Ŧ0.4
C o m p a r a t o r
Ŧ4 Ŧ0.6
t o g g l e v a l u e
Ŧ5
Ŧ0.8
16k 32k 48k 16k 32k 48k
Input code Input code
I n p u t C o d e
C o m p a r a t o r DNL uncalibrated DNL calibrated
o f f s e t
1.5
0.5
1
Fig. 5. Comparator offset calibration.
DNL [LSB]
DNL [LSB]
0.25
0.5
0 0
Ŧ0.5
Ŧ0.25
the reference point. Both the reference point and the first Ŧ1
Ŧ0.5
comparison point are selected externally via a control vec- Ŧ1.5
16k 32k 48k 16k 32k 48k
tor. The comparison result should indicate that the refer- Input code Input code
ence point is lower so the procedure continues by sweeping
the comparison point down in 1/4 LSB steps until the com- Fig. 6. Simulated DNL and INL min/max envelopes with
parator toggles. The comparator offset is calculated at this 13-bit LSB array linearity.
toggle point. Correct toggle point is guaranteed because the
1/8 LSB calibration source is connected with the reference
value only. The measured comparator offset is subtracted clear improvement in static linearity and that the residual
from each cumulative offset resulting in adequate cancella- non-linearity originates from the LSB array. The possibilit-
tion. ies of linearizing the LSB array by extending the calibration
range are discussed. By using the presented calibration, a
DAC with 12-bit intrinsic static linearity, can be improved
4. SIMULATIONS so that 16-bit linearity is achieved.
The effect of the calibration on static linearity was simulated
with MATLAB. 1000 DACs were generated with two dif- 6. REFERENCES
ferent LSB array matching profiles. Matching of the MSB
current sources corresponded to the requirements of a 12-bit [1] G. A. M. Van der Plas, et al., “A 14-bit Intrinsic Accur-
DAC. acy Q2 Random Walk CMOS DAC”, IEEE Journal of
If the matching of the LSB current sources corresponded Solid-State Circuits, pp. 1708-1718, Dec. 1999.
to the linearity requirements of a 13-bit DAC, the calibration [2] D. Mercer, “A 16-b D/A Converter with Increased
improved the DNL yield from 0% to 99.7% and the INL Spurious Free Dynamic Range”, IEEE Journal of Solid-
yield from 0% to 89.2%. If 14-bit matching was defined State Circuits, pp. 1180-1185, Oct. 1994.
for the LSB sources, the improvement of the DNL and INL
yields were from 0% to 100% and from 0% to 97.8%, re- [3] D. W. Groeneveld, H. J. Schouwenaars, H. A. H. Ter-
spectively. meer, and C. A. A. Bastiaansen, “A Self-Calibration
Envelope curves for the minimum and maximum DNL Technique for Monolithic High-Resolution D/A Con-
and INL values of the 13-bit LSB matching are presented in verters”, IEEE Journal of Solid-State Circuits, pp.
Fig. 6. Worst case DNL is calibrated from 1.5 LSB down to 1517-1522, Dec. 1989.
0.7 LSB. INL is improved from 5.5 LSB to 0.75 LSB. [4] Y. Manoli, “A Self-Calibration Method for Fast High-
Resolution A/D and D/A Converters”, IEEE Journal of
5. CONCLUSIONS Solid-State Circuits, pp. 603-608, Jun. 1989.
A digital start-up calibration for current-steering DACs has [5] Y. Cong, R. L. Geiger, “A 1.5V 14b 100MS/s Self-
been proposed. The calibration method does not require Calibrated DAC”, Solid-State Circuits Conference, Di-
additional calibration DAC or A/D converter, but uses a gest of Technical Papers, pp. 128-129, Feb. 2003.
simple comparator for mismatch measurement. The idea [6] E. A. Vittoz, “ The Design of High-Performance Ana-
of skipping over transfer function discontinuations and the log Circuits on Digital CMOS Chips”, IEEE Journal of
use of four redundant MSB sources to achieve full 16-bit Solid-State Circuits, pp. 657-665, Jun. 1985.
input range has been demonstrated. The simulations show
,
A STATISTICAL METHODOLOGY FOR THE DESIGN OF HIGH-PERFORMANCE
CURRENT STEERING DAC’S
ABSTRACT
Random device variations are a key factor limiting the perfor-
mances of high-resolution CMOS current steering D/A convert-
ers. In this paper a novel design methodology based on statistical Y)
modeling of MOS drain current has been developed. This tech-
nique requires firstly an estimation of mean value and autocorre-
lation function of a single stochastic process, which all the pro-
cessldevice variations are lumped in. Then a behavioral model of Figure 1: Error source modeling parameter mismatch.
D/A converters has been developed. Finally the statistical simu-
lation of static performances (DNL and INL) has been carried out
for different DAC architectures.
On the basis of such considerations the following simple
model of the random term will be assumed
1. INTRODUCTION
11 = I& (IT.,L ) y (x,y) (1)
Designing high-performance DAC’s requires a deep understanding
of the relationship between performance and source of technolog- where the term I: accounts for the device operating region (being
ical errors, especially for high-resolution DAC’s to reach a higher /3 a fitting parameter), g (111 L ) takes into account the dependence
effective number of bits. Earlier works have presented several ap- on dimensions and y(z:y ) summarizes all the sources of error de-
proaches facing this problem [l, 21. pending on the device position in the die. Therefore y(z,y) can be
Recently new architectures using segmentation have been sug- considered as a spatial stocliastic process. while 11 is a stochastic
gested which benefits the advantages of both the binary- and the process, assumed to be with zero mean, not only dependent on the
thermometer-coded architecture [3]. However one of the main coordinates ( 2 ;y) but also on bias and device sizes.
lacks in designing high-performance DAC’s is the statistical mod- Referring to a circuit with N devices placed at different posi-
eling of the error sources, which results not adequate to predict tions in the die, Ai sources of error act in the circuit. The statistical
statistical effects on performance. In particular assuming uncor- properties of such a system can be summarized by means of the
related error sources, as done in previous works, not only gives autocorrelation function R,,,,resulting in
unreliable results but also results scarcely useful in reduction mis-
match effect through a suitable layout design. Instead a reduction R,,, ( . % . I . ?/r,
~ ) E ( 7 1 ( r ( Y!
j ~ ,j = . 171 (r.1 Y I ) } i
io = I D + 71. (1)
The non-random term I D represents the usual dc drain current, and
while the term 71 gives raise to some random fluctuations (around
I D )depending on the device position in the wafer, the region of
operation of the device and its dimensions T I ‘ and L .
This work was pnrtially supported by CNR - Prog. Final. MADESS I1 Eq. (4) is of central importance since it relates the covariance of
V-311
0-7803-6685-9/01/$10.000200l IEEE
Shared
Gate 6 2
where ( I , , IC.,.. and IC,, are fitting parameters and T , ~= :r, - x;,,,
T,,= U , - {I, are distances between a pair of devices along :r:- and
y-axis respectively. As a consequence C,,,,predicts a dependence
on ( . I . : y). other thin on voltages and dimensions.
V-3 12
- -
lo
DAC
ARCHITECTURE
OUT
PERFORMANCE
B
+
08-
06.
Figure 5 : Simulation at system level.
03
02.
I .vdd
It can be shown that DNL and INL are related to the vector 77
through the following matrix equations
where
Figure 6: (a) Binary weighted DAC, (b) Thermometer coded DAC.
mances which characterize the behavior of the system, i.e. Differ- being Dnl and In1 matrices independent of the architecture.
ential Non-Linearity (DNL), Integral Non-Linearity (INL), ... Eqs. (13) and (14) establish a linear dependence between the
ii) Define for a given architecture a relationship relating the performances DNL, INL and the random sources 77. Thus it is
DAC output OUT to the vector of all current sources io and a straightforward to derive the covariance matrices of DNL and INL
mathematical model relating the performances cp to the output.
At the end of this process, assuming a linear relationship be- C D V DLN L = J D N L . C,,, . J T n n r ~ , (17)
tween L+Y and io (achieved through linearization in case of non-
linearity), it is straightforward to derive the covariance of the per- C I .L ~I h ' L = J l h L . c,,,. JTA~L. (18)
formances as a function of the covariance of the currents, resulting These equations, whose form is the same as eq. (IO), are the re-
in quired relationships for statistical simulation of DAC architectures.
c;, = s ' c,,,, sT
' (10)
where S is a matrix depending on the DAC architecture. 5. APPLICATION EXAMPLES
4.1. Non-dynamical performances On the basis of the behavioral DAC models previously derived the
statistical simulation of the performance can be carried out.
To derive an input-output relationship relating DAC output to the Before proceeding with such a task it is worth to note that in
vector error sources q ,we define OUT as the vector whose com- general every current source Z D in a DAC is implemented with a
ponents are the 2" output levels and which will be assumed as the set of sub-sources I'D (of LSB value or less) whose total current ex-
output of the DAC. Referring to the architecture shown in Fig. 6(a) actly fits the 7,9 value. In such a way the large number of freedom
for a binary weighted DAC, that in Fig. 6(b) for a thermometer degrees can be used to reduce statistical technological variations.
coded DAC and to a mixed mode DAC which uses both the ar- Thus as the covariance matrix C,,,,refers to the errors q of the cur-
chitectures, a relationship relating the output OUT to the current rents i n , a new matrix C , , J ,taking
~ , into account the errors q' of
vector i~ can be derived. Since all the architectures considered sub-currents );2 has to be defined. Because every current Z D is a
are linear networks, the dependence of output on source currents sum of one or more &,, the vector q is achieved from q' through
will be linear as well. Therefore, it results a linear transformation L, that is
OUT = T iD (11) q = L 77' (19)
being T a matrix dependent on the specific architecture whose ex- so that it results
plicit relationships have been derived for the three architectures G1',
= L C,f,,f LT
' '
mentioned before.
In order to obtain non-dynamical performances DNL and INL, In the simulation examples we assume all the error sources y are
it is useful to derive the vector _10[TT whose components are correlated with the autocorrelation function given by eq. (7). The
the random variables associated with the ideal output levels. By random samples have been generated by a suitable software pro-
combining eqs. ( 1 ) and ( 1 1 ) it results gram according to the distribution chosen. The covariance matrix
behavior derived in such way for a weal; correlation between error
1 0 1'T = T ' q (12) sources is reported in Fig. 7.
V-3 13
To describe the current sources arrangement we use a matrix
notation in which the term Bx (Tx) represents a sub-source related
to the x-th binary (thermometric) source. With such a notation, re-
femng to a 6-bit purely binary DAC, the current sources arrange-
ment shown in the box of Fig. 8 has been chosen. In this case
the results achieved are not satisfactory as clearly appears from
Figs. &(a)and 8(b) reporting the diagonal terms of the covariance
matrices C D N L D , V L and C I , V L
I , W L respectively (normalized to
the variance of LSB), that is the variance of the performances as
function of the input code. As you can see both DNL and INL have 35
a maximum variance corresponding to the half-code transition. 30
Better results have been obtained with a layout solution sug- 25
gested in [I], which takes into consideration mismatch effect. The 0
,
20
arrangement chosen shown in the box of Fig. 9 is such that all
15
the six binary sources have been split into arrays of not contigu-
10
ous unitary sources. As you can see in Fig. 9(a) the maximum in
the variance at half-code disappears, as well as the values of the 5
(b)
variance at the other transitions decrease. Moreover the maximum 0
10 20 30 40 50 60
values result even less than the ones in the case of uncorrelated er- Input Code
ror sources, meaning that correlation between random sources can
be suitably exploited. Also INL improves with such a layout as it Figure 8: Diagonal terms of the covariance matrices
clearly appears from Fig. 9(b). (a) C O\ L D\ L and (b) C , \ L 1 L for a 6-bit DAC (current
Let us now consider a thermometer DAC. In this case the vari- sources are ‘arranged as shown in the box).
ance of DNL ( ~ 6is simply oiSB
. ~ ~ for every transition. With re-
_ _ _ . _ ~
gard to INL it is worth noting that in order to reduce its variance a
B 1 8 6 85 86 8 3 86 B5 86
suitable sequence of current source selection, named “hierarchical 86 8 3 8 6 05 8 6 0 4 8 6 8 5
symmetrical switching” [ 2 ] , has been adopted. The thermometer 16 85 86 81 86 B5 86 82 86
V-3 14
IEEE 2004 CUSTOM INTEGRATED CIRCUITS CONFERENCE
I ...........
1#
Figure 2 Static digital storage correction. eo
The switches which redirect the current either to the output
node or the calibration hardware act as the cascode devices 60
and thus fix the drain voltage of the main current source device 4cI
to be the same, within the matching of the Vgs of the two
cascode switches, in both cases. This can result in a much 20
more accurate calibration. The additional circuitry used for the
calibration is not clocked during normal operation and does
not use power or inject noise into the main signal path. Figure 4 Single vs double cascode
While it is true that output impedance requirements are greatly
reduced for fully differential output configurations, as pointed
out by Luschas in [9], it is important to design the output
switches and their gate voltages so as to keep the output
switches in saturation. This maximizes the attenuation of the
output swing 'seen at the common source nodes of the
differential switches. The small signal attenuation of the
switches is given by the ratio of the device gp to gds. Typical
values of this ratio are in the range of 20.
T
Figure 3(a) Code dependent output impedance
Finite Output Impedance
*e
Vbias
mJ
I lmsb
186 9-5-2
Early work in BiCMOS by Mercer in [16] pointed out that switch element turns on at a different point in the transfer
high-speed operation requires a small and linear source node function and as a result will have a different wave shape on
capacitance for the switches.. A switch unit element will see an node Cs. In figure 6(a) the complementary outputs IA and IB
attenuated output signal on the switch common source node, are shown. Referring to figure 5 , node Cs will have the
and any non-linear back gate capacitance, depicted as C1 in attenuated version of IA when M1 is on and the attenuated
figure 5 , on this node will produce distortion. Tying the switch version of IB when MI is off ( M 2 on ). In figure 6@) we see
and cascode transistors’back gates to the supply reduces non- what the signal on Cs will look like when a switch element is
linear capacitances, but for a large array, the total non-linear near the lower end of the transfer function. Similarly,for figure
capacitance can be significant. The input of a unity-gain level- 6(c) we see the signal on Cs when a switch element is near the
shifting amplifier can be connected to the switch common top of the transfer function.
source node and used to drive the back gate of the switches
and cascode [4]. The non-linear back gate capacitances now In figure 5 , note that the point at which M1 and M2 switch is
see the signal on both plates, thereby bootstrapping the well determined by the crossing point of gate drive signals G1, G2
capacitances and leaving small linear parasitic capacitances. with respect to the value of node Cs. If the relative value of Cs
The amplifier’s dc level shift should set high to minimize the is modulated by the output swing and where in the transfer
switch’s non-linear capacitance. function the switch element is, the actual time point when the
switches change will also be a function of the output swing and
Switch Gate Drive their position in the transfer function. This will result in a
It is very necessary to optimize the waveform, i.e. the crossing signal dependent jitter seen in the output. As indicated in
point of the gate drive signals of the current source switch pair figure 6@) M1 switches from on to off when Cs is at its low
as shown in figure 5 . The circuit that drives the differential point and M1 switches from off to on when Cs is near the high
switches Ml,M2 needs to ensure that the current is always point. For the case shown in figure 6(c) just the opposite
flowing at a constant value. If the voltage at node Cs is happens. The amount of jitter depends on the magnitude of the
allowed to move then some of the current will flow in the signal on Cs and the rise/fall time of the gate drive signals.
parasitic capacitance, C 1, shown in the figure. The best way to
insure that the current flow is constant is to have the gate
signals cross at a point below the nominal DC voltage of the
. common node Cs equal to the Vgs of one of the switch devices
( M1,M2 ) when 1/2 of the current is flowing through the
switch. This minimizes the excursion of the voltage at Cs
during a transition and hopefully the disturbance is symmetric
around the nominal DC value as indicated by the center
waveforms in the figure. An important point is that it is not
necessary to bring the gates of the switch devices any higher
than the voltage on the Cs node when turning off the device
(Vgs = 0). This reduces any feed through of the gate drive
signals to the outputs or the common source node Cs.
time, nSeconds
Figure 7 Switching delay vs output swing
An example case is shown in figure 7, where the normalized
zero crossing point of the differential output voltage at IA,IB
is shown for three cases. The horizontal axis is 5 pSec per
division and the vertical axis spans about 1 mV. The three
curves are for cases where the difference between IA and IB
when the switch flips is -333 mV, 0 V and +333 mV. For these
three cases the node Cs has shifted its nominal value by a total
of 32 mV or approximately 1/20 of the output. We see a shift
in time of 4 pSec, which results fiom a differentialslew rate on
the gate drive signals Gl,G2 of 125 pSecN. This could be a
significant source of error when generating high frequency
outputs.
(c)
Figure 6 Common Source Waveforms Similarities between DAC and Flash ADC
Another source of dynamic error relates back to the fact that a The thermometer coded segments of a switched current DAC
small attenuated amount of the output signal leaks through the are very much analogous to the full parallel flash ADC. The
gds of the differential switch onto node Cs. The amplitude of complexity and hardware of both doubles for each bit of
the signal seen at node Cs is typically about 1/20 that seen at resolution. In the ADC, the distribution of the analog input
outputs IA and IB or 50 mV for a 1V swing at the output. Each signal to the comparators with matched delays is much the
9-5-3 187
same as the collection and combining of the individual unit Digital data pattern dependent noise
current outputs of the DAC. Also, as in the Flash ADC where Gonzales in [lo] observed that noise generated by the data
the delays in the clock distribution to the individual passing through the digital logic portions, specifically the
comparators must be tightly matched, the clock distribution thermometer decode section, of a DAC can cause spurious
network driving the final stage of re-timing latches in the DAC tones and distortion in the analog output. Gong [ 1 11 teaches us
is equally important. that it is possible to concentrate this noise at the clock
frequency Fs or Fs/2. This is accomplished by including a
shadow or mirror data path with a one to one correspondence
to the main data path. This shadow data path is driven by a
data pattern is such a way that for each node in the main data
path that does not change value at a give clock transition the
correspondingnode in the shadow path does. Likewise, when a
node in the main path does change the corresponding shadow
node does not change. This makes the sum total of all nodes
changing at each clock transition constant and independent of
the data pattern.
An example of this technique in an over-sampled switched
current audio DAC is proposed by Rueger in [17]. In this
design a dummy data shift register creates constant local
digital edge activity on the supply, ground and substrate.
NMOS switch devices, driven by full rail swings, are used to
switch the cascoded PMOS current sources. The use of the
dummy data to drive dummy switch devices balances the
Delay = A switching activity injected into the output stage thus
minimizing the demodulation of out-of-band noise into the
jClOCk base band.
Input
Figure 8 Distribution by Propagation Matching A similar notion referred to as modified mismatch shaping
One possible approach to this is propagation delay matching (MMS) is proposed by Shui in [12]. The idea is to set the
[4]illustrated in figure 8. Here if we assume that each cell has number of elements or cells switching per clock period to a
constant. This tums the errors caused by non-ideal element
the same delay A and the delay along the clock distribution
dynamics into a dc offset and energy at Fsl2. Shui assumed an
line from cell 1 to cell n is 61 and the delay along the output oversampling converter, where the maximum output
line is 62 then the sample timing is preserved if61 = S2. bandwidth is reduced. The choice of what fraction of the total
number elements to set the constant to is problematic and the
optimum is a function of the nature of the signals being
$1
converted, however. In any case, the constant can never be set
to more than 1/2 the number of elements. This limits either the
maximum amplitude or the maximum output frequency to only
Clock DAC 1/2 of what it would have been otherwise. Therefore, we
-
Input
-
Output conclude that, for a Nyquist rate converter, to make use of this
constant element switching concept we would need twice as
many elements.
Data Dependent clock loading
A rather important special case of data pattern dependent
-
timing error comes from the varying load seen by the final
clock driver which drives the last re-timing latch of the DAC.
The basic issue is that all latches to some extent present a load
to the circuit, which drives the clock input, that depends on
Delay A whether the content ( state ) of the latch is changing or not.
Figure 9 Distribution by Constant Wavefront Matching Given the finite strength of the driving circuit, the rise time of
Binary tree distribution structures are often used to match the clock waveform will be a function of the number of the
these delays as well as done by Schafferer in [15]. This results multiplicity of latches connected to this common clock driver
in a constant wavefront as illustrated in figure 9. The clock which are changing their state at a given time. This results in a
distribution tree is arranged to have equal lengths ffom the shift in time of the output samples which is a function of the
driver to each cell. Likewise the output collection tree is rate of change on the waveform and thus gives rise to odd
arranged with equal length ffom each cell to the output pad(s). order distortions mainly third order.
The clock tree delay does not need to match the output tree
delay. An example of this effect is shown in figure 10. The time when
the rising edge of the clock signal crosses mid supply (1.25 V)
is plotted for the case of a single latch when the input data is
The physical placement of the unit cells in the layout is an
important consideration and geometric shuffling of the not changing and when the input data is changing. This
placement is often used to breakup any linear gradients in the simulation shows a 2 pSec difference. This effect is magnified
when a large number of latches are driven by the same
cell delay A ( figure 8,9 ) that might be present.
188 9-5-4
common clock driver and is proportional to the number of differential outputs. Switch M1 is switched off for 1/2 of the
latches which change their state. In the case of thermometer clock period while the DAC current switches (IDAC) change
coded data the number of data bits changing is proportional to and then back on after the currents have settled. Resistor R1
the rate of change of the output waveform. provides a load while M 1 is off. As well, there is switch M2 to
1.8 short the output to ground, through resistor R2, when the other
I I I I I I I switch is off thus the return-to-zero operation. This effectively
reduces the timing skew between the various DAC switches.
There is however a loss of 1/2 of the signal amplitude, 6&,
due to the return-to-zero output.
I I I I I I I
-20 -15 -10 -5 0 5 10 15 20
time, pSeconcla
Figure 10 Data dependent clock delay Figure 12 Return to Zero Concept, output waveform
We can get a workable solution by taking the shadow or mirror
Return to zero switching can reduce the distortion from digital
data paths concept of [ l l ] and combining it with what we
data noise induced timing errors, but for very high sample
concluded from [121 and realize that by doubling the number
rates if the output does not completely settle in each half of the
of latches by simply adding the second as a mirror path for cycle then the history effect or inter-symbol interference is not
each original latch and driving the mirror latch in such a way eliminated and can result in signal dependent distortions. It is
as to cause it to change state only when the main data latch important to note that this scheme is not totally free of signal
does not. One way this mirror data can be generated is shown
,
dependent timing errors. The turn on and turn off points for
in figure 11. By combining the main data signal with a clock
M1 and M 2 will depend on the signal levels seen at the node
signal at Fs/2, or 1/2 the main clock rate, with an exclusive OR
OUT. It is difficult to tell if the reported SFDR results for this
gate the mirror data signal is created such that it changes only
method are really any better than those reported in [4]because
when the main data does not. By doubling the number of
both seem to be limited to about -75 dBc for output
latches we have doubled the load on the clock driver but it is
fkquencies above 25 MHz, which seems the measurement
now independent of the incoming data pattern.
limit of most spectrum analyzers
LFp
I
4
Fclock
"\ %Driver
k:
clock cycle. When constant switching techniques are used the
distortion or noise is concentrated as a tone at the sampling
frequency.
189
By using four switches instead of the normal two, we are in
effect interleaving two return to zero switches. The Quad switching like this incorporates some of the good points
configuration of the quad switch is shown in figure 13. There of both RZ switching and ordinary differential switching, is
are four switch devices M1, M2, M3 and M 4 which share a suitable for high sample rates, and reduces transition-
single common source connection Cs. The unit element current dependent noise. A drawback of quad switching is an increase
Im is supplied to node Cs as in the ordinary differential switch. in complexity, four gate signals need to be generated and the
Only one of the four switches is on at any given time as increased dynamic power consumption due to the fact that one
indicated by the switching waveforms of figure 14. pair of the four switches each cycle.
Conclusions
,I lmsb Seven major contributors to errors and distortion in modem
switched current Digital-to-Analog converters have been
discussed. Static device matching can be addressed either
though statistical averaging or calibration. One or more
cascodes can be included, along with insuring that the output
switches remain in saturation, to reduce the effect of output
impedance variation. The importance of gate drive signals was
explored. Much like the flash ADC, clock distribution is a key
factor. Digital data pattern generated noise needs to be
addressed and the effect on clock jitter can be a major source
of distortion. Return-to-zero switching can be employed to re-
time the output sampling time. The use of a quad switch and
constant data activity switching techniques can shift spurious
outputs to the sampling frequency.
u
[6] Vandenbussche, J. et al.”A 14 b 150 Msample/s update rate 42 random
walk CMOS DAC”, ISSCC 1999 Digest of Technical Papers, 15-17 Feb.
G3 1999 Page(s): 146-14
[7] Groeneveld D.W.J., Schouwenaan H.J., Tenneer H.A.H., Bastiaansen
C.A.A.; “A self-calibration technique for monolithic high-resolution D/A
t L I converters” IEEE Joumal of Solid-state Cicuits,Volume: 24 , Issue: 6 ,
G4
I
I \ U
I Dec. 1989 Pages:1517 - 1522
[8] A.Van den Bosch, et al., “SFDR Bandwidth Limitations for High Speed
High Resolution Current Steering CMOS D/A Converters,” Proc. ICECS,
pp. 1193-1196,1999.
‘“B
IA
190 9-5-6
IDDQ TESTING OF A CMOS 10-BIT CHARGE SCALING
DIGITAL - TO - ANALOG CONVERTER
A Thesis
in
by
Srinivas Rao Aluri
Bachelor of Engineering, Osmania University, Hyderabad, India, 2000
December 2003
Acknowledgements
I would like to dedicate my work to my parents, Mr. and Mrs. A.Sreedhar Rao
and my brother Seshu, for their constant prayers and encouragement throughout my life.
I am very grateful to my advisor Dr. A. Srivastava for his guidance, patience and
encouragement have helped me to get a deep insight in the field of VLSI design.
I would like to thank Dr. P. K. Ajmera and Dr. Martin Feldman for being a part of
my committee.
I would like to thank my friend Miss. Chandra Srinivasan for her constant love
I take this opportunity to thank my friends Anand, Uday, Vijay and Harish for
their help and encouragement at times I needed them. I would also like to thank all my
friends here who made my stay at LSU an enjoyable and a memorable one.
Last of all I thank GOD for keeping me in good health and spirits throughout
my stay at LSU.
ii
Table of Contents
Acknowledgements...........................................................................................................ii
Abstract ............................................................................................................................x
iii
3.6.1 BICS in Normal Mode ......................................80
3.6.2 BICS in Test Mode ...........................................82
3.6.3 Detailed Analysis of BICS ...............................83
3.7 Layout, Simulation and Timing Diagrams for BICS ...................84
3.7.1 Current Mirror Circuit.......................................84
3.7.2 Current Differential Amplifier..........................84
3.7.3 BICS..................................................................87
3.8 Fault Detection, Simulation and Testing .....................................87
3.8.1 Fault-Injection Transistor...................................89
Bibliography: ..................................................................................................................123
Vita...................................................................................................................................136
iv
List of Tables
4.1 Theoretical and measured IDDQ for different fault types........................................119
v
List of Figures
2.9b Equivalent circuit with MSB =1, and all other bits set to zero ............................23
2.12 Layout of a unit capacitance made of poly1 and poly2 used in the design .........28
2.13 Layout of the capacitor array using unit capacitor configuration ........................29
2.14 Layout showing the use of dummy capacitors to match the capacitors
present at the corner of the capacitor array..........................................................31
vi
2.19 Layout of an operational amplifier design of Fig. 2.16 .......................................40
2.21 Input and Output response of the amplifier circuit of Fig. 2.16 ..........................42
2.24 The CMOS operational amplifier as a unity gain amplifier (follower) ...............48
2.30 Sample and hold response of circuit of Fig. 2.26 obtained from post
layout SPICE simulations ....................................................................................58
vii
3.7a Gate-oxide-short (GOS) in a MOSFET ...............................................................74
3.7b Equivalent circuit model. RS is the effective resistance of the short. B models
the rectifying behavior of new current path introduced by the defect .................74
3.8 Bridging fault causing IDDQRB drop and a path to the ground .............................76
3.18 CMOS MUX circuit with defect 3 introduced using a FIT ..................................95
3.19 CMOS MUX circuit with defect 4 introduced using a FIT ..................................96
4.1 CMOS chip layout of a 10-bit charge scaling DAC with four
fault injection transistors distributed across the chip............................................98
4.2 MOS chip layout of a 10-bit charge scaling DAC including BICS within a
padframe of 2.25mm × 2.25mm size ....................................................................99
4.3 Microchip photograph of 10-bit charge scaling DAC and BICS for
IDDQ testing...........................................................................................................101
4.4 Simulated and measured characteristics of a 10-bit charge scaling DAC ...........102
viii
4.7 Voltage gain response of op-amp with fault introduced .......................................105
4.8 Transfer function of op-amp with fault (VE1) induced..........................................106
4.9 Gain versus frequency response of the CMOS opamp circuit with
fault introduced .....................................................................................................107
4.10 Simulated output response of the multiplexer circuit of Fig 3.18
without defect........................................................................................................108
4.11 Simulated output response of the multiplexer circuit of Fig. 3.18
with fault activated................................................................................................109
4.12 Simulated BICS output of the circuit of Fig 3.16 when Error-signal-1
for defect-1 is activated.........................................................................................110
4.13 Simulated BICS output of the circuit of Fig 3.17 when Error-signal-2
for defect-2 is activated ........................................................................................112
4.14 Simulated BICS output of the circuit of Fig. 3.18 when Error-signal-3
for defect-3 is activated.........................................................................................113
4.15 Simulated BICS output of the circuit of Fig.3.19 when Error-signal-4
for defect-4 is activated.........................................................................................114
4.16 CMOS circuit diagram of 10-bit charge scaling DAC with four
fault injection transistors distributed across different parts of the circuit.............115
4.17 Simulated BICS output with defects induced using fault injection transistors.....116
4.18 HP 1660CS logic analyzer test results on a fabricated CMOS 10-bit charge
scaling DAC showing the performance of BICS in normal and test modes.........118
ix
Abstract
This work presents an effective built-in current sensor (BICS), which has a very
small impact on the performance of the circuit under test (CUT). The proposed BICS
works in two-modes the normal mode and the test mode. In the normal mode the BICS is
isolated from the CUT due to which there is no performance degradation of the CUT. In
the testing mode, our BICS detects the abnormal current caused by permanent
manufacturing defects. Further more our BICS can also distinguish the type of defect
induced (Gate-source short, source-drain short and drain-gate short). Our BICS requires
neither an external voltage source nor current source. Hence the BICS requires less area
and is more efficient than the conventional current sensors. The circuit under test is a 10-
x
Chapter 1
Introduction
Conventional logic testing applied in digital circuits can detect faults, which can
cause logic errors [1]. It, however, cannot detect several physical defects such as the gate-
oxide shorts, floating gates and the bridging faults. These physical defects do not cause
any logical error at the time of testing but can manifest into a fault at an early stage of the
circuit life. Thus, circuits should also be tested for physical defects apart from logic
errors. Though methods exist for testing digital circuits but testing of analog circuits is
still a problem. Mixed signal circuits are even more difficult to test. A general and
efficient solution for testing mixed-signal integrated circuits is still not available.
Functional test approach applied to test the functionality of analog and mixed-
signal integrated circuits is based on empirical development of a test set [2, 3]. This
approach needs a reasonably large number of sample circuits for collecting the test data.
The approach also does not have any inherent test metric to measure the achievement of a
test goal. Design for testability (DFT) is another widely used method [4]. Oscillation test
strategy is based on the DFT technique [4, 5], which gives good fault coverage and does
not require any test vectors. In this method, the complex analog circuit is partitioned into
blocks. During the test mode, each building block is converted into a circuit that
oscillates. The oscillation frequency, fOSC of each building block can be expressed as a
function of its components or performances. The oscillation test method allows removing
the analog test vector generator and output evaluators, and consequently reduces the test
1
complexity, area overhead, and test cost. However, the method suffers from performance
degradation in complex integrated circuits since it is not usually possible to divide the
circuit into the fundamental blocks. Built in self-test (BIST) method is based on
measuring the output data and calculating the performance of the system using an on-chip
circuits by incorporating all or some of the testing circuitry on the silicon. An important
component of a mixed-signal BIST is a precision analog signal generator required for on-
chip stimulation. While the area overhead is kept to a minimum, these generators should
frequency and amplitude. This method also suffers from performance degradation and
The steady state or quiescent current (IDDQ) testing of CMOS integrated circuits is
known to be very efficient for improving test quality [8, 9]. The test methodology based
on the observation of the quiescent current on power supply lines allows a good coverage
of physical defects such as gate-oxide shorts, floating gates and bridging faults. These
defects are neither well modeled by the classical fault models, nor detectable by
conventional logic tests. In addition, IDDQ testing can be used as a reliability predictor due
to its ability to detect defects that do not yet involve faulty circuit behavior, but could be
transformed into functional failures at an early stage of circuit life. Thus, IDDQ testing
became a powerful complement to the conventional logic testing. In analog circuits, the
quiescent current, termed as IPS may be in the order of µA’s or mA’s [10]. Under the
fault conditions, the normal values of IPS may be increased, decreased or generally
2
distorted. Thus, fault detection can be accomplished by monitoring the IPS current
In this thesis, a simple built-in current sensor (BICS) is presented, which provides
a digital output for supply current monitoring and testing in mixed-signal circuits. The
output proportional to the IDDQ (IPS) current of the circuit under test. BICS is inserted in
series with the power supply or the ground of the CUT to detect abnormal IDDQ current in
the integrated circuit [5] as shown in Fig. 1. The new BICS requires less area and has less
neither an external voltage reference nor a current reference since the reference is
generated on chip. It only requires two extra control pins, which control the mode of
operation of the BICS and one output pin. The proposed BICS has been designed to test a
IDDQ’s definition is the current that flows in the CMOS circuit under quiescent (steady
state) operating condition. Any current above the quiescent current would indicate the
presence of physical defects in the circuit. Figure. 2 shows how an IDDQ test can identify
physical defects. Q4 has a defect that causes a resistive path between its gate and source.
When VIN is logic ‘1’ level, the gate and the source of Q4 are held at ground by Q2, which
prevents the flow of any IDD current [20]. When VIN switches to a logic ‘0’ level, Q4’s
gate is pulled high by Q1. This allows the current to flow from VDD through the defective
path to ground.
3
VDD
PMOS
BLO C K
IN P UT S
O UT P UT
NMOS
BLO C K
P AS S /F AIL
BIC S
4
1.2 Reliability Benefits Derived from IDDQ
IDDQ testing has been shown to identify gate oxide shorts (GOS). Many of these gate
oxide shorts do not initially cause a functional failure but over the time they can
Inc. (FMI) conducted a life test study on IDDQ failures for ASIC’s, recording leakage
current instability [20]. It was theorized that any instability could be used as a leading
indicator for predicting component reliability. After 48 hours of life testing, the ASIC
components passed all burn-in, hot and cold functional testing, and parametric testing but
failed for IDDQ. The circuit under test (ASIC components) had IDDQ values ranging from 0
µA to greater than 100 µA [20]. This experiment was repeated and verified on similar
devices. Failing IDDQ does not necessarily result in non-functional behavior. However,
data is available, which confirms a significant number of IDDQ failures will result in
reliability problem [20]. Life test studies conducted at FMI have shown a statistically
significant number of IDDQ failures have become inoperative over time [20]. This life test
In the following section a brief review of the selected works on IDDQ testing are
• Maly and Patyra’s design [13]: The main idea is to make use of a differential
amplifier with a reference voltage at the non-inverting node and the current-
related voltage at the inverting node. Two clocks φ1 and φ2 are used to control
when to sample the data. The importance of minimizing the impact of BICS on
the performance of the CUT is taken into consideration. Hence a lateral NPN BJT
5
VD D VD D
ID D
Q1 Q3
Vin Vo u t
Q2 Q4
D efec t
6
• is used to supply a large current with an allowable drop [13,18]. Due to the
exponential characteristic of the BJT, this design [13] also enhances the IDDQ
detectability even when the abnormal current is little more than steady state
• Favalli et al. design [14]: In this design, each gate requires two extra transistors
and the whole circuit requires other transistor (see Fig. 3 of [14]). All these
transistors are used to convert analog faults to stuck-at faults. A selection line is
used to switch between normal mode and test mode. The design is easy to
implement since it uses only digital circuits. However, the area overhead is quite
• Miura and Kinoshita’s design [15]: The design consists of a V-I translator, a level
translator and a integrator circuit. The V-I translator transforms the current of a
given CUT to a corresponding voltage. The level translator transforms the output
voltage of the V-I translator to an appropriate logic value. The integrator circuit
produces a faulty signal when the faulty logic generated by the level translator
lasts for certain time intervals. This BICS itself can be tested by applying an
external current source to the special ground pin NGND (Fig.1 of [15]).
• Shen et al. design [16]: The design is similar to the comparator circuitry (the
two-phase clock CLK1 and CLK2 so that the transient current can be bypassed. A
diode is used to limit the voltage drop on the sensing device. The comparator
7
circuit compares the externally generated IREF with IDDQ from the CUT. Through
proper sampling in the steady state, a PASS/FAIL flag is then detected from the
output circuit. This scheme may still degrade the circuit since the cut-off voltage
0.6V of the diode may be too high when compared to the low level output voltage
VOL = 0.5V. The design also requires external current reference and a two-phase
clock generator.
voltage drop on BICS is further reduced. The current supplied by the CUT passes
through the current sensing transistor Ts. This current is compared with the
reference current. Since the transistor, Ts operates in the linear region, less current
can be provided for a given device dimension. Thus, even if the voltage stability
is also needed. [See U.S. patent 5057774, Oct. 15, 1991 by S.C. Verhelsts, E.
integrated monolithic digital circuits” referred in J.J. Tang, K.J. Lee and B.D. Liu,
“ A practical current Sensing Technique for IDDQ Testing,” IEEE Trans. on Very
large Scale Integration (VLSI) Systems, vol.3, No.2, June 1995, pp.302-310].
Methods have been developed to test integrated circuits based on dynamic supply
current (IDD) measurement [21], on-line power dissipation measurement and IDDQ testing
[22] and on-chip transient current measurements [23-26]. Current sensors have been
proposed for analogue applications [22]. In [22], the sensor design is based on a series
voltage regulator, in which a series transistor is connected between the supply and the
8
(CUT). One drawback is the area required to realize this serial transistor since it has to
are discussed.
Chapter 2 explains the basic structure and operation of a 10-bit charge scaling digital to
current sensor. The mechanism of fault simulation and fault detection in a 10-bit charge
Chapter 4 describes the simulation result and design considerations of each module of
10-bit charge scaling DAC. Simulations results of each module of the DAC are included.
It also describes the simulation results for each module of the built-in current sensor.
Finally, a description of the abnormal current behavior and fault detection in the DAC is
explained and simulation results are included. Experimental results of the fabricated
device are presented, compared with simulations. Results are also compared with the
published work.
Chapter 5 provides a summary of the work presented and scope for future work.
The MOS model parameters used for design is presented in Appendix A. The
9
Chapter 2
The ability to convert digital signals to analog and vice versa is very important in
are applied to the input of the DAC to create from a reference voltage an analog output
signal that represents the respective digital word. In this conversion process, an N-bit
digital word is mapped into a single analog voltage. Typically, the output of the DAC is a
Where VOUT is the analog voltage output. VREF is the reference voltage. F is the fraction
defined by the input word, D, that is N bits wide. The number of input combinations
represented by the input word D is related to the number of bits in the word by
The maximum analog output voltage for any DAC is limited by the value of some
reference voltage VREF. If the input is an N-bit word, then the value of the fraction, F, can
be determined by,
D
F= (2.3)
2N
Figure. 2.1 shows a conceptual block diagram of a DAC converter. The inputs are a
digital word of N-bits (b1, b2, b3… bN) and a reference voltage, VREF. The voltage output,
10
VREF
MSB
DN-1
DN-2
Vout
Input word, D
(N bits wide)
DN-3 Digital-to-analog
Converter
(DAC)
D1
D0
LSB
11
Where K is a scaling factor and the digital word D is given by
b1 b2 b3 bN
D= 1
+ 2 + 3 + ……+ N (2.5)
2 2 2 2
N is the total number of bits of the digital word, and bi is the ith coefficient and is either 0
or 1. Thus, the output of a DAC can be expressed by combining Eqs. 2.4 and 2.5 to get
b1 b2 b3 bN
VOUT = K VREF ( 1
+ 2 + 3 + ……+ N ) (2.6)
2 2 2 2
The basic architecture of a DAC without a sample and hold circuit at the output is
shown in Fig. 2.2. The various blocks are a voltage reference, which can be supplied
externally, binary switches, a scaling network, and an output amplifier. The voltage
reference, binary switches, and scaling network convert the digital word as either a
voltage or current signal, and the output amplifier converts this signal to a voltage signal
The following are some of the important static and dynamic performance
The DNL gives a measure of how well a DAC can generate uniform analog LSB
Where ‘n’ is the number corresponding to the digital input transition. DNL is illustrated
in Fig. 2.3.
12
Vref KVrefD Vout = KDVref
Voltage Scaling Output
Reference Network Amplifier
Binary
Switches
b1 b2 b3 bN
13
0.8
0.6
0.4
0.2
LSB's
0
0 200 400 600 800 1000
-0.2
-0.4
-0.6
-0.8
Digital Input Code
(INL). It is defined as the difference between the data converter output values and a
14
reference straight line drawn through the first and last output values as shown in Fig. 2.4.
INL defines the linearity of the overall transfer curve and can be described as follows.
INLn = (output value for the input code, n) – (output value of the reference line at the
point).
• Offset
Offset of the DAC is defined as the amount of shift in the transfer characteristics when
the digital input code D = 0. This shift is similar to the offset voltage for an operational
• Gain Error
Gain Error is defined as the difference between the ideal slope and the actual slope of the
• Dynamic Range
Dynamic range is defined as the ratio of the largest analog output value (Full Scale (FS))
to the smallest analog output value. The dynamic range in decibels is given by,
For our design, which is a 10bit, charge-scaling DAC, the dynamic range is 60.19 dB.
• Resolution
It is described as the smallest change in the analog output with respect to the value of the
15
0.9
0.8
0.7
0.6
0.5
LSB's
0.4
0.3
0.2
0.1
0
0 200 400 600 800 1000
Input Code
16
Actual
Ideal
Normalized Analog output
17
Ideal
VREF 2V
Resolution(N) = Log2 ( ) = Log2( ) = 10 bits. (2.8)
1LSB 1.9mV
A wide variety of of DAC architectures exist, ranging from very simple to complex.
Each of course, hast its own merits. There are primarily three architectures of DAC
namely-
• Resistor String
• Current Steering
• Charge Scaling
18
• Voltage Division
In this architecture, the analog output voltage is divided uniformly among the
resistor string as shown in Fig. 2.7 [27]. Depending on the input digital word, the
switches shown close or open if the input is a ‘high’ or ‘low’ voltage, respectively. The
analog output is simply the voltage division of the resistors at the selected tap. The value
of the voltage at the tap associated with the ith resistor is given by [27]
(i ).VREF
Vi = N
, for i = 0, 1, 2… 2N-1 (2.9)
2
This architecture typically results in good accuracy, provided that no output current is
required and that the values of the resistors are within the specified error tolerance of the
converter. Another problem with this architecture is the balance between the area and
power dissipation. So this architecture is not suited for high resolution DAC’s.
19
VREF
R2 N S2 N-1
V2 N-1
R2 N-1 S2 N-2
V2 N-2
S2
VO UT
V2
R2 S1
V1
R1 S0
V0
R0
• Current Steering
The current steering based DAC architecture is shown in Fig. 2.8 [27]. This DAC
architecture uses current through out the conversion known as current steering. This type
of DAC requires precision current sources that are summed in various fashions. Since
there are no current sources generating iOUT when all the digital inputs are zero, the MSB,
20
D2N-2, is offset by two index positions instead of one. The binary signal controls whether
or not the current sources are connected to either iout or GND. The output current iout has
the range of
One advantage of the current steering DAC’s is the high-current drive inherent in the
system. Of course, the precision needed to generate high resolutions is dependent on how
well the current sources can be matched or the degree to which they can be made binary
weighted. Another problem associated with this architecture is the error due to the
switching.
• Charge Scaling
A very popular architecture used in the CMOS technology is the charge scaling
DAC and is shown in Fig. 2.9(a). In this architecture, a parallel array of the binary-
any value. After initially being discharged, the digital signal switches each capacitor to
either VREF or ground (GND) causing the output voltage, VOUT, to be a function of the
voltage division between the capacitors. Since the capacitor array totals 2NC, if the MSB
is ‘high’ and the remaining bits are ‘low’, then a voltage divider occurs between the MSB
capacitor and the rest of the array. The analog voltage, VOUT becomes
21
D2 N--2 D2 N--3 D2 N-4 D1 D0
iO UT
I I I I I
22
VOUT
2N-1C 2N-2C 4C 2C C C
VREF
DN-1 DN-2 D2 D1 D0
2N-1 C
VREF VOUT
2N-1 C
Figure 2.9 (b): Equivalent circuit with MSB =1, and all other bits set to zero.
VREF
VOUT = , (2.11)
2
23
which confirms the fact that the MSB changes the output of a DAC by ½ VREF. Figure
2.9(b) shows the equivalent circuit under this condition. Therefore, the value of VOUT for
N −1
VOUT = ∑D
k =0
k 2k-N . VREF where k = 0, 1…N-1. (2.12)
The 10-bit DAC used in our design uses charge scaling DAC. The unit capacitance in the
DAC is 10fF. The reference voltage used is 2V, VSS is –2.5V and VDD is +2.5V.
The basic circuit diagram of a 10-bit charge-scaling DAC is shown in the Fig.
2.10. This circuit converts the 10-bit digital input word to a respective analog voltage
depending on the capacitive network. The various blocks associated with the DAC are
multiplexer switches to which the digital word is given. Figure 2.11 shows the integrated
capacitance network, multiplexer switches and amplifier part of the Fig 2.10. Initially the
input digital word is given to a multiplexer circuitry. Depending on the logic value of
each bit of the word, the multiplexer chooses the particular voltage to which the capacitor
is to be charged. If the input bit in the digital word is logic ‘0’ then the multiplexer
chooses the input which is connected to the ‘GND’ and the capacitor is charged to ‘GND’
and if the input bit in the digital word is logic ‘1’ then the capacitor is charged to VREF.
The capacitor at the end of the network is used as a ‘terminating capacitor’. Depending
on the capacitors, which are charged to different voltages based on the input digital
24
STORAGE UNITY-GAIN
TG SWITCH
CAPACITOR BUFFER
CAPACITOR ARRAY ARCHITECTURE OPAMP
-
A=1
- +
+ Vo
+
CH -
512C 256C 128C 64C 32C 16C 8 C 4C 2C C C
O utput of Multiplexer
25
VREF
Control Signal
VDD
M8 M7
W/ L =45/3
W/ L =90/3
W/ L =90/3
M6
W/ L =90/3 W/ L =90/3
512C 256C 128C 64C 32C 16C 8C 4C 2C C C
CC =30FF
- Vin M1 M2 Vout
26
W/ L =9.6/3 M9 M3 M4
W/ L =90/3
M5
W/ L =45/3
W/ L =45/3
V SS
Figure. 2.11: Integrated capacitance network, multiplexer switches and amplifier part of the circuit of Fig. 2.10
The capacitor at the end of the network is used as a ‘terminating capacitor’. Depending
on the capacitors, which are charged to different voltages based on the input digital
words, the effective resultant analog voltage is calculated for the respective digital
combination. This analog voltage is passed through the OPAMP and through the S/H
circuit and appears as analog voltage. Thus, the digital to analog conversion is performed.
The DAC consists of several blocks and their design is explained in following section.
The capacitor architecture in the DAC is being drawn using two poly layers poly1
and poly2. The unit capacitance used in the capacitor array is 10fF. Figure 2.12 shows the
layout of the unit capacitor used in the design. This design considers unit capacitor
percent from the desired value. Because of this, it is difficult to produce high accuracy
capacitors in a standard CMOS process as well as integrated circuits, which rely on the
accuracy of a single capacitance value. If, instead, capacitance ratios are used, the relative
error is cancelled since it is the ratio of the capacitance that is taken in to consideration
but not the single capacitance value alone. Figure 2.13 shows the layout of the capacitor
array using unit capacitor configuration. The array is surrounded with dummy capacitors
and guarded by the guard ring to cancel out the effect of parasitics.
The capacitors, which are present at the end of the arrays, do not have the
surrounding capacitors to cancel out the relative error. To take care of these capacitors
27
Figure 2.12: Layout of a unit capacitance made of poly1 and poly2 used in the design.
28
Figure 2.13: Layout of the capacitor array using unit capacitor configuration.
29
dummy capacitors are added to the array [31]. Figure 2.14 shows the use of dummy
capacitors in the capacitor array layout. The substrate noise present in the substrate can
be coupled to the capacitor through its parasitic capacitor and any voltage variation
present is also coupled to other components of the chip. To avoid this coupling the
capacitor array is shielded from the substrate with N-well under it and connecting it to a
quiet DC potential [31]. The guard rings are used in the layout around the capacitor array
to its input in terms of a known mathematical operation. This is usually achieved by using
active or passive elements such as resistors and capacitors in integrators and differentiator
circuits. A typical operational amplifier is characterized by a high open loop gain, high
bandwidth, a very high input impedance, low output impedance and an ability to amplify
differential mode signals to a large extent and at the same time, severely attenuate
common mode signals. The amplifier in the 10-bit DAC (Fig.2.10) is realized by one
such operational amplifier. Figure 2.15 shows the block diagram of an operational
amplifier [30]. An OPAMP normally consists of four main functional blocks. First is the
input differential gain stage that amplifies the voltage difference between the input
terminals, independently of their average or common mode voltage. Most of the critical
parameters of the OPAMP like the input noise, common mode rejection ratio (CMRR)
and common mode input range (CMIR) are decided by this stage. The differential to
single-ended conversion stage follows the differential amplifier and is responsible for
30
Dummy Capacitor
Figure 2.14: Layout showing the use of dummy capacitors to match the capacitors
present at the corner of the capacitor array.
31
V+ Second
+ + DC + +
Input + Differential
Differential to Gain V0
Single-Ended Level
V- Amplifier - -
- Conversion Stage
32
Shift
the second gain stage properly, a level-shifting block is introduced after the differential to
single-end conversion stage. Finally, additional gain is obtained in the second gain stage.
as shown in Fig. 2.16. This circuit configuration provides a good voltage gain, a good
common mode range and good output swing. Before the analysis of the OPAMP is done,
some of the basic principles behind the working of MOS transistors are reviewed. The
transistors (M1 & M2) with their sources tied together. The differential pair is biased by
current mirrors, which act as the active load too. Two current mirrors: a p-MOS current
mirror (M7 & M8) and an n-MOS current mirror (M3 & M4) are used instead of just one
in order to increase the common mode rejection ratio (CMRR) of the differential pair.
The p-MOS current mirror serves as a constant current source and the n-MOS mirror,
which sinks current, acts as an active load across which the first stage output is taken,
Current mirrors are used extensively in MOS analog circuits both as biasing elements and
as active loads to obtain high AC voltage gain [28, 29]. Enhancement-mode transistors
remain in saturation when the gate is tied to the drain, as the drain-to-source voltage
33
VDD
M8 M7
W/L=45/3
W/L=90/3
W/L=90/3
M6
+ Vin
W/L=90/3 W/L=90/3
- Vin CC =30FF
M1 M2 Vout
M9 M3 M4
W/L=9.6/3 W/L=90/3
M5
W/L=45/3
W/L=45/3
VSS
34
(VDS) is now always going to be greater than the gate-to-source voltage (VGS) due to the
Based on Eq. (2.13), constant current sources are obtained through current mirrors
transistor. Figures 2.17 and 2.18 shows the p-MOS and n-MOS current mirrors design. A
p-MOS mirror serves as a current source while the n-MOS acts as a current sink. The
voltage developed across the diode-connected transistor is applied to the gate and source
of the second transistor, which provides a constant output current. Since both the
transistors have the same gate to source voltage, the currents, which when both transistors
are in the saturation region of operation, are governed by the following equation 2.14
assuming matched transistors. The current ratio IOUT / IREF is determined by the aspect
ratio of the transistors. The reference current that was used in the design is 100µA. The
For identical sized transistors, the ratio is unity, which means that the output current
mirrors the input current. Because the physical channel length that is achieved can vary
substantially due to etching variations, the accurate current ratios usually results when
devices of the same channel length are used, and the ratio of currents is set by the channel
width. For the n-MOS current mirror design shown in Fig 2.18,
35
VDD
M8 M7
(45/3) (90/3)
I0 = 200µA
IREF = 100µA
IREF = 100µA
M3 M4
45/3 45/3
VSS
36
IOUT / IREF = (W4/L4) / (W3/L3) (2.16)
The reference current that is applied to the current mirror is obtained by means of
an active resistor. A resistor can be obtained by simply connecting the gate of a MOSFET
to its drain as shown in Fig 2.17 and 2.18. This connection forces the MOSFET to
Where all the symbols have their usual meanings. Since the gate is connected to the
drain, the current IDS is now controlled directly by VDS and therefore the channel
transconductance becomes the channel conductance. The small signal resistance is given
by
where gm is the transconductance of the MOS transistor. It is defined as the ratio of the
change in drain current to a change in the applied gate and is described by the following
equation.
or gm = 2 β ID . (2.20)
It is to be noted that the transconductance of a MOS increases as the square root of the
drain current. Therefore, MOS amplifiers need several stages to achieve large gains due
37
The operational amplifier designed in this work is shown in Fig. 2.16. The small
Where ISS is the differential amplifier bias current and Vthp is the threshold voltage
of the pmos transistors forming the differential pair. The differential amplifier needs to be
biased by a constant current source, which is provided by the 100µA current source. The
same current is supplied to the two stages of the operational amplifier by the p-channel
current mirrors M8, M7, M6 which provide the bias current for the two stages. In the first
stage i.e. the differential amplifier stage not only is the differential amplification
accomplished but also the differential to single ended conversion done. Thus, the output
is taken only from one of the drains of the transistors. The n-channel devices M3 and M4,
which are the load for the p-channel devices, also aid in the single ended conversions.
The second stage provides a level shift for the output of the differential amplifier stage
and it also provides the additional gain. It is once again biased by a current source, which
is also used to maximize the gain of the second stage. To get a high gain with reasonable
high output resistance the minimum channel length used is 3µm and maximum width of
the transistor used is 90µm. Transistor M5 is critical to the frequency response, is biased
at ID5 = 200µA and has (W/L)5=(W/L)max = 30. The input pair is biased at –ID7 = 200µA.
To avoid input offset voltage transistors M3 and M4 are dimensioned according to [28]
(W L) − I D 6 200µA W 1 W
5
= = = 1 → = = 15 (2.22)
2 * (W ) I D7 200µA L 3, 4 2 L 5
L 3, 4
38
Therefore, the W = 45µm for the transistors M3 and M4. To obtain the bias current of
100µA a MOS resistor is used with appropriate value of width. (which is the MOSFET
simulating resistors). Large W/L ratios for the transistors in the operational amplifier are
obtained by using the following technique. The four transistors are connected in such a
way that the effective W/L ratio is four times the W/L ratio of each transistor. The
and is well characterized. The simplicity, modularity and predictability of the device
The physical layout of the amplifier was made using the L-EDIT 8.20 and the
‘spice’ netlist is extracted including parasitic capacitances. The layout of the amplifier is
shown in the Fig.2.19. Figure. 2.20 shows the transfer characteristics obtained from DC
sweep analysis. The output offset voltage is approximately 33µV. Figure 2.21 shows the
obtained at the output of the OPAMP with peak-to-peak amplitude of 2v, giving a gain of
4000. Figure 2.22 shows the frequency response characteristics. Figure 2.22(a) shows the
amplitude versus frequency behavior. The 3dB bandwidth of the amplifier obtained is
approximately 100 kHz and 3dB gain is 77. Figure 2.22(b) shows the phase versus.
frequency response. The phase noise margin as shown in Fig.2.22 (b) is 200. The
39
Figure 2.19: Layout of an operational amplifier design of Fig. 2.16.
40
Output,V
2.5
2
1.5
1
0.5
0
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1
-0.5 0 0.1 0.2 0.3 0.4 0.5 0.6
-1 Input,V
-1.5
-2
-2.5
DC Offset Voltage = 33 µV
41
Gain = 4000
2.5
Output
2 Input
1.5
1
0.5
Voltage(V)
0
-0.5 0 100 200 300 400 500
-1
-1.5
-2
-2.5
Time(us)
Figure 2.21: Input and Output response of the amplifier circuit of Fig. 2.16.
42
90 Gain = 80dB - 3dB = 77dB
80
70
60
Gain (dB)
50
3 dB Band Width
40
= 100 KHz
30
20
10
0
1 10 100 1000 10000 100000 1E+06 1E+07 1E+08
f (Hz)
43
50
45
40
Phase (degrees) 35
30
25
20
15
10 Phase margin = 200
5
0
10 100 1K 10K 100K 1M 10M 1G
frequency (Hz)
90
80
70
60
Gain (dB)
50
40
30
20
10
0
10 100 1K 10K 100K 1M 10M 1G
frequency (Hz)
44
2.3.6 Unity Follower
A unity follower is basically an amplifier with a gain of unity. It is used mainly as
a buffer amplifier in order to increase the current driving capability of the amplifier stage.
An ideal unity follower would exhibit infinite input impedance, zero output impedance,
large bandwidth and unity gain. The gain-bandwidth product of the amplifier is known,
as it’s figure-of-merit, and is a constant for any given amplifier. It is usually determined
From basic OPAMP theory, we know that the gain of an OPAMP in its non-
Rf
V0 = (1 + ) Vi (2.23)
R
Rf and R are the feedback and series resistances of the OPAMP and Vi and V0 are the
input and output voltages of the OPAMP. The gain of the OPAMP is determined by the
Rf
Av = (1 + ) (2.24)
R
If Rf is zero, then the gain of the amplifier is unity. The input to be buffered is applied to
the non-inverting terminal of the unity follower, and the output connected to the inverting
terminal of the OPAMP in the feedback configuration. So, as the signal increases in
strength at the non-inverting terminal, the signal at the inverting end increases too thus
forcing the output to follow the input. Only the differential stage of the amplifier
discussed was used to realize the unity follower buffer as shown in Fig 2.24. The drain of
transistor M5 is coupled back to the gate of the input transistor M1. This connection
45
Rf
- Vo
+
Vi
Vo = (1 + Rf ) Vi
R
46
introduces feedback to the inverting terminal of the OPAMP, thus, putting it in the unity
follower configuration.
output of the DAC during conversion. The characteristics of the S/H amplifier are crucial
to system accuracy and the reliability of the analog data. As its name indicates, a S/H
amplifier has two modes of operation, programmed by a digital control input. In the
sample mode, the output follows the input, usually with a gain of unity. When the mode
input switches to hold, the output of the S/H amplifier ideally retains the last value it had
when the command to hold was given, and it retains that value until the mode input
switches back to sample. At this time, the output ideally jumps to the input value and
follows the input until the next hold command is given. Figure 2.25 is a block schematic
gate switch (TG-switch), a storage capacitor CH and a unity gain follower. Each of these
mechanism is as follows. The TG-switch is operated by the VCONTROL signal and is closed
during the sample interval and open during hold. So, during sample, the circuit is
connected to promote rapid charging of the storage capacitor, CH and during hold, the
capacitor, CH is disconnected from its charging source and ideally retains its charge. The
held by the storage capacitor. The unity gain buffer is used at the output to avoid the large
overshoot, which might occur, on the output when the input changes quickly.
47
VDD
M8 M7
W/L=45/3
W/L=90/3
W/L=90/3
M6
+ Vin
W/L=90/3 W/L=90/3
CC =30FF
- Vin M1 M2 Vout
W/L=9.6/3 M9 M3 M4
W/L=90/3
M5
W/L=45/3
W/L=45/3
VSS
Figure.2.24: The CMOS operational amplifier as a unity gain amplifier (unity follower).
48
STORAGE UNITY-GAIN
TG SWITCH BUFFER
CAPACITOR
Mp -
SOFT NODE
A= 1
+
+ Vo
+
Mn CH -
VH
VCONTROL -
49
2.3.8 The Transmission Gate Switch
The input switch to the S/H amplifier is controlled by a digital signal generated
from an external control and synchronizing logic to represent the two modes of operation
- the sample and hold. The switch should ideally offer zero resistance to the signal when
it is closed (sample) and infinite resistance when open (hold). Moreover, when closed, it
should exhibit the same properties irrespective of whether the capacitor is charging or
discharging. An ideal switch also takes zero time to turn on or off. A CMOS transmission
transistor. The transmission gate switch can be made to turn ON or OFF for either
polarity of the mode control signal by connecting a simple inverter between the gates of
the transistors. Since the transmission gate has both an n-type and a p-type devices,
The expression for the charging time constant through a CMOS transistor gate can be
expressed as [32]
1
τn = CH [ + Req] (2.26)
βn (VDD − Vtn)
1
τp = CH [ + Req] (2.27)
βp (VDD − Vtp )
where CH is the hold capacitor. The term Req represents the equivalent resistance of the
transmission gate which remains a constant during charging and discharging. From the
above two equations, if the threshold voltages of the n-MOS and p-MOS transistors are
made equal and the aspect ratios of them are so adjusted such that βn = βp, the time to
charge and discharge through the transmission gate would be equal, thus giving
50
symmetrical response. Since the transition from ON to OFF and vice versa is much faster
now with the use of a transmission gate switch, two important terminologies with
reference to S/H amplifiers are defined. Acquisition Time is the time required by the
output of the S/H to reach its final value, within a specified error band, after the sample
command has been given. Aperture (Delay) Time is the time required for the switch to
open fully after the hold command is given. The held voltage is, in effect delayed by this
interval and the hold command should therefore be advanced by this amount for precise
timing. Needless to say, both these times are reduced significantly by the use of a
The storage capacitor limits the slew rate in the sample mode and determines the
‘droop’ in the hold mode of operation. The slew rate is the rate at which the voltage
across the capacitor can change with respect to time and is entirely a function of the input
signal frequency. The equivalent circuit of the S/H amplifier during sample is that of a
low-pass filter with the series resistance of the filter consisting of the TG-switch
resistance when closed, and the storage capacitor CH . For the voltage of the capacitor to
follow the input signal fairly well, the RC time constant of the filter should be close to the
time period of the input signal. The value of the storage capacitor to be used is therefore a
function of the input signal frequency. The other consideration for the value of the
storage capacitor is the droop rate. ‘Droop’ is the gradual drop in the ‘held’ voltage by
the capacitor with time, during the hold period. Obviously, this introduces errors in the
digital-to-analog conversion process, as the voltage level at any time after the instant it
was sampled would be different from the level at which it was sampled.
51
The storage capacitor CH was implemented using the Poly1 and Poly2 layers in
standard CMOS process. The parallel plate capacitance used in the design is 12 pF.
Figure 2.26 shows the S/H CMOS circuit. The layout of the sample-and-hold CMOS
circuit of Fig 2.26 is shown in Fig. 2.27. The post-layout SPICE simulations were
performed. Figure 2.28 shows the post-layout simulated sample and hold response of the
circuit of Fig 2.26. The input to the circuit is a 4V p-p sine wave and the control voltage
given to the circuit is 5V p-p pulse. When the pulse is HIGH, the circuit samples the
input and when the pulse is LOW, the circuit holds its previous state. A microvolt signal
was applied to the input of the OPAMP and a control signal Vcontrol was applied to the
sample-and-hold circuit as shown in Fig. 2.29. The amplified signal from the OPAMP
was reproduced by the unity follower with a total offset of about 33µV. Figure 2.30
shows the circuit behavior of the circuit of Fig. 2.29 obtained from post-layout
simulations.
Figure 2.31 shows the layout of a 10-bit DAC. The 10-bit charge scaling DAC is
tested by giving various combinations of digital input words and the respective analog
output voltage is obtained. The referenve voltage used in the design is 2.0V. The 10-bit
charge scaling DAC has about 1024 digital word combinations and is quantized within
the reference voltage of 2.0V with a step of 1.9mV. This is obtained as follows
The reference voltage used is 2V. The least significant change in the output value is
2
LSB = = 1.9mV. (2.28)
1024
52
Figure. 2.32(a) and (b) show the DAC output characteristics, for all combinations of the
53
STORAGE
TG SWITCH
CAPACITOR UNITY GAIN BUFFER
VDD
M8 M7
W/LN=4.5/1.6
W/L=45/3 W/L=90/3
W/L P =4.5/1.6 W/L=90/3
M6
W/L =4.5/1.6
Mp W/L=90/3 W/L=90/3
54
CC =30FF
+ -V M1 M2 Vout
Input of S/H Mn CH in
W/L =4.5/1.6 M9 M3 M4
VH W/L=9.6/3
W/L=90/3
M5
W/L=45/3
- W/L=45/3
(VCONTROL)
VS S
VO UT VCO NTRO L
55
2.0V
0V
-2.0V
V(Input)
4.0V
-4.0V
V1(Vcontrol)
4.0V
0V
-4.0V
0s 50us 100us 150us 200us 250us 300us 350us 400us
V(Output)
Time
56
OPERATIONAL AMPLIFIER STORAGE UNITY GAIN BUFFER
TG SWITCH
VDD CAPACITOR VDD
M8 M7 M8 M7
W/L=45/3 W/L=45/3
W/L=90/3 W/L=90/3
W/L=90/3
M6 W/L=90/3
M6
+ Vin
W/L=90/3 W/L=90/3 Mp
- Vin CC =30FF W/L=90/3 W/L=90/3
M1 M2 CC =30FF
+ - Vin M1 M2 Vout
M9 M3 M4 Mn CH
W/L=9.6/3 W/L=90/3
M5 VH W/L=9.6/3 M9 M3 M4
57
W/L=45/3 W/L=90/3
W/L=45/3 M5
W/L=45/3
- W/L=45/3
VSS (VCONTROL)
VSS
Figure 2.29: Schematic block diagram showing op-amp, TG-switch, storage capacitor and unity gain buffer.
2.0V
0V
-2.0V
I/P of OPAMP
2.0V
0V
-2.0V
O/P of OPAMP
4.0V
-4.0V
V1(Vcontrol)
4.0V
0V
-4.0V
0s 50us 100us 150us 200us 250us 300us 350us 400us
V(Output)
Time
Figure 2.30: Sample-and-hold response of the circuit of Fig. 2.26 obtained from post-
layout SPICE simulations.
Note. Input signal peak-to-peak voltage is 4V.
58
Multiplexer
TG - Switch
59
2
1.5
0.5
0
0 1000 2000 3000 4000 5000 6000
Time(us)
(a)
2.5
2
1.5
1
0.5
0
0 256 512 768 1024
Digital Input Code
(b)
Figure.2.32 DAC output response for all (0000000000 – 1111111111) combinations of
the input digital word.
60
Chapter 3
This chapter focuses on IDDQ testing using built-in current sensors (BICS), the
design and implementation of the BICS in a 10-bit charge scaling DAC, the fault
simulation and detection methodology. It also discusses previously proposed schemes for
IDDQ testing and the important physical faults commonly seen in the design of integrated
circuits. Simulations and design considerations for the BICS are also discussed.
IDDQ testing of CMOS ICs is shown very efficient for improving test quality. The test
methodology based on the observation of quiescent current on power supply lines allows
a good coverage of physical defects such as gate oxide shorts, floating gates and bridging
faults, which are not very well modeled by the classic fault models, or undetectable by
conventional logic tests [1]. In addition, IDDQ testing can be used as a reliability predictor
due to its ability to detect defects that do not yet involve faulty circuit behavior, but could
be transformed into functional failures at an early stage of circuit life. Due to obvious
quality and reliability improvements, this approach became powerful complement to the
and efficient technique for mixed-signal testing, where fault detection of analog parts
requires the precise measure of the IDDQ. In analog circuits, the quiescent current, termed
as IPS, may in the order of µAs or even mAs. Under fault conditions, the normal values of
IPS may be increased or decreased or generally distorted. Thus, fault detection can be
accomplished by monitoring the IPS current fluctuations. Figure 3.1 shows the fault free
IDDQ current in the quiescent state, which is about 1mA and shooting to 3mA when fault
61
is injected in the CUT [19]. Elevated IDDQ does not necessarily result in nonfunctional
behavior. However, data are available confirming that IDDQ failures will result in
reliability problems [33]. Considerable impact can be made towards achieving higher
Built-in current sensors (BICS) have speed and resolution enhancements over
off-chip current sensors, mainly because the large transient currents in the output drivers
are by-passed and less parasitic are encountered. On chip current testing is both time-
efficient and sensitive. Moreover, on-chip current tests can also be used as an on-line
testing tool, and is important when components are to be used in high reliability systems.
For high speed and high sensitivity, unaffected by large pad currents, a fast built-in
current testing circuit is desired [34]. In the present work, a simple design of a built-in
current sensor is presented to detect bridging faults in a 10-bit charge scaling DAC. A
novel method has been introduced for the fault injection to simulate physical defects
present in a chip.
IDDQ measurements require analog circuitry that can ideally measure current
below 1 µA in the range 10 KHz- 33 MHz [35]. Different methodologies exist for IDDQ
testing. IDDQ testing can be classified in two groups, 1) external IDDQ testing and 2)
internal IDDQ testing. External IDDQ testing monitors power supply current through the
power pins of the integrated circuit package while internal IDDQ testing monitors power
62
+ 2.5v
VERROR - SIGNAL
-2.5v
1mA
IREF
3mA IDEF
IDDQ
1mA
10US 20US 30US 40US 50US 60US 70US 80US
Time
63
3.2.1 External IDDQ Testing
The simplest form of hardware for testing is the automatic testing equipment
(ATE) precision measurement unit. It can be connected to the CUT’s power pin and used
to measure IDDQ. This strategy is acceptable if the number of measurements is less than
20; otherwise test time becomes expensive [1]. Figure 3.2 shows off-chip instrumentation
schematic. Figure 3.2 (a) illustrates an approach for an off-chip IDDQ instrumentation
[33]. The tester is connected in series with the VDD line of the CUT. CDD is the total
capacitance at the VDD node and includes that due to the IC itself and any capacitance
added by the tester and instrumentation circuit. The tester measures dVo/dt during the
quiescent time and can estimate IDDQ if CDD is known using [1].
Figure 3.2 (b) shows another schematic for off-chip IDDQ current measurement [35]. An
off-chip sensor monitoring the power supply current is a simple implementation of IDDQ
testing and widely used in production testing. The transistor Q1 is ON only during
transient when the CUT is drawing large currents. To filter the high impedance noise at
high frequencies, a small capacitor C1 in the range of 2-2.5nF is added in between the
sense circuit and the CUT [37]. Once transients are settled, the Q1 is OFF and capacitor
C1 supplies the static current to the CUT. IDDQ is measured by the voltage drop across the
transistor Q1.
Off-chip current measurement technique has the ability to detect vast majority of
manufacturing defects, including those that are not detected by the traditional stuck-at
fault testing [34]. However, off-chip measurement techniques have speed and sensitivity
64
V0
TESTER
VDD IDD
1pF
CUT
CDD
.
Figure 3.2 (a): Off-chip IDDQ current measurement using an automatic test equipment.
n-MOS
Sense Circuit
To power
CUT Q1 supply
C1
C1
To
sample &
hold
65
limitations [38]. Low-current resolution is critical in detecting defects such as floating
gates, which do not cause large abnormal currents. Off-chip measurements may not
detect this small current due to its sensitivity limitations. Several other factors can
degrade off-chip IDDQ instrumentation. 1) All testers have current probes, which offer
significant capacitive loading at the power supply, causing a large voltage drop across it,
and lack in DC accuracy [1,20]. 2) The test board exists in a noisy electrical environment
and long leads are used and electromagnetic fields are high [35]. Current measurement is
slow and susceptible to static noise in the power supply bus. Considerable noise is
therefore introduced into the measurement. 3) Above all, the major portion of the IDDQ
current in CMOS VLSI chips is generated at the output pad circuits, and abnormal IDDQ
current is overshadowed by the output currents [36]. Owing to these limitations with off-
chip current measurements, the built-in current sensor is a preferred approach in many
applications. It can be integrated in to the CMOS design to test for physical faults in the
circuit.
The effectiveness of IDDQ testing can be enhanced if built-in current sensors are
applied on chip to monitor defect –related abnormal currents in the power supply buses
[34]. This testing technique applies on-chip current sensors that detect abnormal power
bus currents and overcomes the limitations of the off-chip IDDQ current measurements.
Essentially, this technique adds a BIC sensor in series with VDD or GND lines of the
circuit under test. A series of input stimuli is applied to the device under test while
monitoring the current of the power supply (VDD) or ground (GND) terminals in the
quiescent state conditions after the inputs have changed and prior to the next input change
66
[35]. Figure 3.3 shows the block diagram of the IDDQ testing with BICS. The many
advantages of BICS over the off-chip current testing or the ATE, include: reduction of
test equipment cost, increase of testing rate, improvement of the detectability and high
static circuit should be negligibly small. However, in some cases, due to charge presence
in a gate oxide or latch-up, the sub-threshold current may be large enough to become an
essential component of IDDQ . The BICS can be designed to detect this current also.
In CMOS technology, the most commonly observed physical failures are bridges,
opens, stuck-at-faults and gate oxide shorts (GOS). These defects create indeterminate
logic levels at the defect site [1]. Very large-scale integrated circuits processing defects
cause shorts or break in one or more of the different conductive levels of the device [36].
We briefly discuss these physical defects that cause an increase in the quiescent current.
Figure 3.4 shows a 2-input NAND open circuit defect. Logic gate inputs that are
unconnected or floating inputs are usually in high impedance or floating node-state and
cause elevated IDDQ [34]. In Fig. 3.4, node VN is in the floating node-state. For an open
defect, a floating gate may assume a voltage because of parasitic capacitances and cause
the transistor to be partially conducting [37]. Hence, a single floating gate may not cause
a logical malfunction. It may cause only additional circuit delay and abnormal bus current
[34]. In Fig. 3.4, when the node voltage (VN), reaches a steady state value, then the output
voltage correspondingly exhibits a logically stuck behavior and this output value can be
67
PMOS
BLOCK
INPUTS
OUTPUT
NMOS
CUT BLOCK
PASS/FAIL
BICS
68
VDD
VA
Q1 Q2
V0
Q3
VN
Q4
69
weak or strong logic voltage. Open faults, however, may cause only a small rise in IDDQ
current, which the off-chip current sensor may not detect because of its low-resolution
[1]. It can be detected using BIC sensors. An open source or open drain terminal in a
transistor may also cause additional power-bus current for certain input states. In this
The short circuit faults in very large-scale integrated circuits are popularly termed
as bridging faults. When IDDQ measurements are used, a bridge is detected if the two nets,
which compromise it, have opposite logic values in the fault-free circuit [37] and are
connected by a bridge due to the introduction of the fault in the circuit. Bridging faults
can appear either at the logical output of a gate or at the transistor nodes internal to a
gate. Inter-gate bridges between the outputs of independent logic gates can also occur.
Bridging fault could be between the following nodes 1) drain and source, 2) drain and
gate, 3) source and gate, and 4) bulk and gate. Examples of bridging fault are shown in
Figs. 3.5 and 3.6, respectively. Figure 3.5 shows example of possible drain to source
bridging faults in an inverter chain in the form of low resistance bridges (R1, R2 and R4).
charge scaling DAC at different conducting levels. We have introduced faults in the 10-
bit charge scaling DAC by using “fault-injection transistors” instead of hard metal shorts
invented in our group [19]. The introduction of a fault via the “fault-injection”
70
VDD
R3 R4
R1
V1 Vo
R2
VS S
Path from VDD to ground
71
VDD
V0
VA
Bridge 1: Drain-gate
VB
Bridge 2 : Gate-source
72
transistor enables the 10-bit DAC to function fault-free under the normal conditions. The
faults considered include source-drain bridge, drain-gate bridge and source-gate bridge.
Bridging defect cannot be modeled by the stuck-at model approach, since a bridge often
does not behave as a permanent stuck node to a logic value [37]. IDDQ testing using BICS
principle physical reasons for GOS are the breakdown of the gate oxide and the
manufacturing spot defects in lithography and processes on the active area and
polysilicon masks [38]. Figure 3.7 illustrates the circuit level gate oxide short defect
model [37]. These defects can be seen as short-circuits between the gate electrode and the
conducting channel of the device through SiO2 . GOS short causes an undesirable current
injection in to the channel [36, 38]. This current injection forces a substantial increase in
the quiescent current. The diode-resistor combination could be used to model the
rectifying behavior of the new current path introduced by the defect [38].
These defects are unlikely to produce logical errors, but cause important deviation
IDDQ’s definition is the level of power supply current in a CMOS circuit when all
the nodes are in a quiescent state. Static CMOS circuits use very little power and at stand-
by or quiescent state, it draws practically negligible leakage current [35]. In steady state,
there should not be a current path between VDD and GND path. Ideally, in a static CMOS
circuit, quiescent current should be zero except for associated p-n junction leakage
73
Gate
Source Drain
n+ n+
p - Si
(a)
S D
Rs
(b)
74
currents. Any abnormal elevation of current should indicate presence of defects. To
assure low stand-by power consumption, many CMOS integrated circuit manufacturers
Figure 3.8 shows how an IDDQ test can identify defects. The current in static
CMOS is not constant during transient [40]. When an output transition occurs, a peak of
IDDQ current is observed. This peak is due to charging and discharging of the load
capacitance at the output circuit and corresponds to the short circuit. When the transition
is completed, the circuit is in the quiescent state. IDDQ is very sensitive to physical faults
in the circuit. In mixed-signal CMOS circuits such as data converters. IDDQ is around
faults. Two nodes connected by a bridge must be driven to opposite logic levels under
fault-free conditions for bridging fault to occur. In Fig.3.8, a typical bridge is one
between the node VO and VDD. To detect this defect, input pattern must drive the node
VO1 to the logic low value (‘0’), as this node is assumed to be bridged with the power rail.
Thus, a path from power to ground appears allowing the existence of an abnormal high
IDDQ current. IDDQ value is directly dependent on the resistance offered by the conducting
path and hence on the size of the transistors in the conducting path. The presence of the
physical fault causing the high abnormal current can be effectively detected by IDDQ
testing using BICS. A set of realistic bridges have been modeled between adjacent
metal lines in a 10-bit charge scaling DAC at three different (conducting levels), to
75
VDD
(IDDQ RB ), V
RB
V1 V01 V02
Figure 3.8: Bridging fault causing IDDQRB drop and a path to the ground.
76
examine the effect on the value of IDDQ and detect the presence of the fault using the
BICS.
A simple design of a BIC sensor built into the 10-bit charge scaling DAC is
presented using the current mode design. It determines whether the circuit quiescent
current is below or above a threshold level. Previously proposed schemes and the
characteristics required for a good BICS are discussed briefly in this section.
Different BICS schemes have been proposed for detection of the abnormal IDDQ
current and the physical faults commonly observed. While most BICS designs
concentrate on mere detection of the fault, some can detect the location of the fault as
well [41]. The entire design is divided into n sub blocks (SB) where n equals the number
of outputs. The divided SB’s are checked individually through their corresponding output
and a faulty area is easily detected by observing the outputs. The performance impact of a
BICS on a circuit under test (CUT) is the key issue to be considered when designing
BICS. Insertion of the BIC sensor between CUT and GND involves series voltages, and
these voltages could degrade the performance of the CUT [40,11]. A large number of
earlier BICS are based on voltage amplifiers such as differential amplifiers or sense
amplifiers. The stability of the BICS is limited in this case since the quiescent point (Q-
point) of an amplifier may not be stable and can vary with the change of dc supply
voltage, VDD. The detection time and hardware overhead is increased due to the extra
77
To overcome problems of slow detecting time, resolution, instability of the BICS
and large impact on the CUT performance, the current-mode circuit design approach has
been adopted using a single power supply. In this work, simple design of a BICS
employing current mirrors and current differential amplifier has been proposed. It has
minimum area over head in the chip and no impact on over all performance.
3. The design should be simple and compact to minimize the additional area
Figure 3.9 shows the CMOS circuit diagram of the built-in current sensor. It
consists of a current differential amplifier (M2, M3), two current mirror pairs (M1, M2 and
M3, M4) and an inverter. The n-MOS current mirror (M1, M2) is used to mirror the current
from the constant current source which is used as the reference current IREF for the BICS.
The current mirror (M3, M4) is used to mirror the difference current (IDEF-IREF) to the
current inverter, which acts as a current comparator. The differential pair (M2, M3)
calculates the difference current between the reference current IREF and the defective
current IDEF from the CUT. The W/L size of the n-MOS current mirrors (M1, M2) is set to
27/1.6 and (M3, M4) is set to 72/1.6. Therefore ID3 = IDEF-IREF. The constant reference
current is set to approximately the same value as the quiescent state current when the
78
VDD
VDD
18/1.6 VDD
M5 CUT
6/1.6
M6 VDD
VSS
I DEF
I REF EXT VSS 150/1.6
IDEF - IREF OUTPUT
I REF
27/1.6 ( PASS/ FAIL )
27/1.6
60/1.6 25/1.6
M2 M3
M0 M1 M4
72/1.6
72/1.6
VENABLE VSS
VSS
79
CUT is fault free. In the present design, the reference current, IREF is set to 1mA. The
output inverter buffer has an aspect ratio ((W/L)P / (W/L)N) of 6/1 to counter capacitive
parasitics at the output node and detect the presence of the physical fault through the
The proposed scheme operates in two modes: the normal mode and the test mode.
The mode of operation is controlled by ‘VENABLE’ signal applied to the gate of transistor
M0. The W/L size of the M0 is 60/1.6. This enables the 10-bit charge scaling DAC, which
is the CUT to operate as fault-free in the normal mode of operation. Further explanation
Figure 3.10 explains the basic structure of the BIC sensor connected between
CUT and GND in IDDQ testing. The BICS is inserted in series with GND or VSS line of
the circuit under test. The proposed BICS works in two modes: the normal mode and the
test mode. The mode of operation is decided by the VENABLE signal. In the normal mode
(VENABLE = ‘1’), the BICS is isolated from the CUT. In the test mode (VENABLE = ‘0’), the
quiescent current from the CUT is diverted in to the BICS and compared with reference
During the normal operation, the signal ‘VENABLE’ is at logic ‘1’ and all the IDD
turned on. Therefore, the n-MOS current mirrors have no effect on dynamic current. It
follows that the BICS’s output is not affected by the dynamic current. Thus, in the normal
mode, the BICS is totally isolated from the 10-bit DAC (CUT). Since in normal mode the
80
VDD
PMOS CUT
BLOCK
VIN
Output
NMOS
BLOCK
VDD
18/1.6 VDD
M5
M 6 6/1.6
VSS
IDEF
IREF EXT OUTPUT
VSS
IDEF - IREF PASS/FAIL
IREF
27/1.6
72/1.6 150/25
60/1.6
27/1.6 M 2 M3
M0 M1 M4
72/1.6
VENABLE
VSS
81
current coming from the CUT is same as the reference current the difference current IDEF
– IREF becomes negligible and the output of the BICS is at logic ‘0’. In the normal mode,
Since the BICS is inserted in series with GND line of the CUT, it causes a voltage
drop and large capacitance between the CUT and the substrate. These effects cause
performance degradation and ground level shift. To reduce these extra undesirable
effects, an extra pin EXT is added to the proposed BICS. Pin EXT is connected to the
drain of transistor M0. In the test mode, it is left floating. In the normal mode, EXT gets
connected to logic ‘0’. In the normal mode, since the EXT pin is grounded by passing the
BICS, the disturbance of the ground level shift during normal operation of the circuit
never happens. Therefore, there is no impact on the performance of the DAC, while the
During the test mode, the ‘VENABLE’ signal is at logic ‘0’. The IDDQ current from
the CUT is diverted by the BICS and the n-MOS current mirror pair replicates the
reference current to the current differential amplifier which assigned a value nearly same
as the fault-free current. This mirrored reference current is compared with defective
current IDEF current coming from the CUT. The output of the current comparator, which
drop of VDS across the transistor M4. In the test mode, the difference current is large
which turns-ON M4 heavily and forces its output node pulled-down to logic ‘0’ and is
detected as PASS/FAIL output ‘1’, indicating presence of defects in CUT. In the testing
82
mode, EXT pin is floating. The ‘VENABLE’ signal is connected to GND and M0 is off. The
timing diagram and detailed analysis of BICS are explained in the following sections.
The current differential amplifier and the current mirror are most important parts
of the proposed BICS. Performance of the current mirror greatly affects the BICS’s
ability to detect abnormal current due to physical defects. The current mirror has a
property that, in a constant current stage, the reference current in one branch of the circuit
is mirrored in the other branch [22]. The current differential amplifier on the other hand
receives reference current IREF one input, and the defective current IDEF from CUT as the
other input. The differential amplifier calculates the difference between the two currents.
The other n-MOS current mirror mirrors the difference current calculated by the current
differential amplifier to the inverter (M4, M6) which acts as comparator (Fig.3.10). The
output of the current comparator is used as the PASS/FAIL flag. If the IDDQ is greater
than the reference current, we presume there are defects within the functional circuit. If
the IDEF is less than the IREF, we assume that the functional circuit is free from physical
defects that induce abnormal IDDQ current. Functionally, in inverter, ID = IREF – IDEF. If
IDEF > IREF, then the PASS/FAIL output shows a logic ‘1’. Conversely if IDEF < IREF, a
logic ‘0’ will appear at the PASS/FAIL output. Input impedance of the inverter is very
large [23]. Because of the high input impedance of the current comparator and the
utilization of the current mirror along with current differential amplifier, even a very
small difference between IDEF and IREF can be distinguished. That means the BICS can
achieve a high resolution. Several SPICE simulations were performed to determine the
83
3.7 Layout, Simulation and Timing Diagrams for BICS
transistor as shown in Fig. 3.11. The voltage developed across the diode-connected
transistor provides the constant current output. The current ratio IOUT / IREF is determined
by the aspect ratio of the transistors in the current mirror circuit design. Both n-MOS
transistors have the same W/L ratio of 27/1.6. The output current of the current mirror
In our design,
The current differential amplifier used in our design is the most important part of
the BICS. The current differential amplifier calculates the difference between the
reference current IREF and defective current IDEF. Figure 3.12 shows the circuit diagram of
differential amplifier. The currents IREF (i1) and IDEF (i2) are the input currents of the
amplifier. The difference current (IDEF – IREF or i2 – i1) is calculated and mirrored to
output of the current differential amplifier. The BICS design comprises of two n-MOS.
current mirrors and a current differential amplifier. One n-MOS current mirror provides
the reference current as one input to the differential amplifier and the other input current
comes from the CUT. The other n-MOS current mirror replicates the difference current to
84
IREF
IOUT
M1 M2
85
VDD
i2
i1 i1 i2 - i1 i2 - i1
M1
M2 M3 i2 - i1
27/1.6 M4
27/1.6
72/1.6
72/1.6
VSS
86
the output inverter which acts as a current comparator. The proper design of the current
mirrors is most crucial for working of the BICS. The input impedance of the current
The current differential amplifier finds applications in both low-power and high-speed
circuit design.
3.7.3 BICS
Figure 3.9 shows the circuit diagram of the BICS. It comprises of two current
mode and 2) test mode. The test signal is applied to an n-MOS transistor, M0 (W/L =
60/1.6), which decides the mode of operation. When the test signal is ‘0’, the BICS is in
the test mode. When the test signal is at logic ‘1’, the BICS is isolated from the CUT and
its output is at logic ‘0’. The output inverter buffer has an aspect ratio ((W/L)P/(W/L)N) of
6/1. Figure 3.13 shows the layout of the BICS of the circuit shown in the Fig. 3.9.
The primary reason for a fault is a defect in the integrated circuit. A manufacturing defect
causes unacceptable discrepancy between its expected performance at circuit design and
actual IC performance after physical realization [34]. A defect may be any spot of
missing or extra material that may occur in any integrated circuit layer.
Two nodes are connected if there is at least one path of conducting transistors
87
IREF IDEF
VSS
BICS
O/P
VENABLE
VDD
EXT
88
between them. If the two nodes are at opposite potentials under fault-free conditions, a
conducting path between them will increase the IDDQ current due to fault in the circuit.
After transient switching, each node in a digital circuit is one of the following four states
1. VDD state: This state occurs when the node is connected to VDD.
2. GND state: This state occurs when the node is connected to GND.
3. Z state: The high-impedance state occurs when the node is neither VDD nor GND
connected.
4. X state: This state occurs when the node is both VDD-connected and GND-connected
[34].
The ‘X’ state should never occur in fault-free CMOS integrated circuits. Many
defects cause an X state to occur in CMOS integrated circuits. Thus, we can view testing
as a way to detect the X state, which causes detectable abnormal steady state current.
Bridging faults have been induced in the DAC at various conducting levels using a fault-
injection transistor (FIT), discussed further ahead, which cause abnormal elevation of the
In this work, four bridging faults have been placed in the 10-bit DAC design using fault-
injection n-MOS transistors. Activating the fault-injection transistor activates the fault.
The use of a fault-injection transistor for the fault simulation prevents permanent damage
to the 10-bit DAC by introduction of a physical metal short. This enables the operation of
the DAC without any performance degradation in the normal mode. Figure 3.14 (a)
shows the fault-injection transistor. To create an internal bridging fault, the fault-injection
89
ME) is connected to VDD, a low resistance path is created between its drain and source
nodes and a path from VDD to GND is formed. In the Fig. 3.14(b), an internal bridging
fault is created in the CMOS inverter between the drain and source nodes using the fault-
injection transistor. Logic ‘0’ is applied at the input of the inverter. Therefore, the output
of the inverter is at logic ‘1’ or VDD. When the logic ‘1’ is applied to the gate (VE) of the
n-MOS fault-injection transistor (ME), it turns on. This causes a low resistance path
between the output of the inverter and the VSS. This gives rise to an excessive IDDQ
current as a path from VDD to GND is created, which can be detected by the BICS.
Figure 3.15 shows the layout of a 10-bit charge scaling DAC with BICS. The area
of the DAC alone is 692 × 502 µm2. The entire area of the CUT along with BICS is 692 ×
516 µm2. Therefore the BICS occupies only 242 × 40 µm2 of the entire chip area. Four
defects have been introduced using fault-injection transistors. The n-MOS fault-injection
transistor (ME) is designed for W/L equal to 4.5/1.6. The fault-injection transistors are
activated externally using ERROR signals VE1, VE2, VE3 and VE4, respectively. Error
signal VE1 is applied to the gate of the fault-injection transistor in defect 1, which forms a
short between the source and drain in the operational amplifier circuit shown in Fig. 3.16.
Error signal VE2 is applied to the gate of the fault-injection transistor in defect 2, which
forms a short between the drain and gate in the unity gain buffer circuit shown in Fig.
3.17 for S/H. Error signal VE3 is applied to the gate of the fault-injection transistor in
defect 3, which forms a short between the gate and the source in the multiplexer circuit
shown in Fig. 3.18. Error signal VE4 is applied to the gate of the fault-injection transistor
in defect 4, which forms a short between the gate and the substrate in the multiplexer
90
VE
G
ME
D S
(4.5/1.6)
VDD
VI V0
D
(4.5/1.6)
ME VE
G
VSS FIT
Figure. 3.14 (b) Fault-injection transistor between drain and source nodes of a CMOS
inverter.
91
Defect 4
Defect 3
Defect 1
Defect 2
Figure 3.15: Layout of a 10-bit DAC with BICS showing the defects induced in the CUT
using fault-injection transistors.
92
VDD
M8 M7
W/L=45/3 W/L=90/3
W/L=90/3
M6
+ Vin
W/L=90/3 W/L=90/3
- Vin CC =30FF
M1 M2 Vout
W/L=4.5/1.6 W/L=9.6/3
M9 M3 M4
VE1 W/L=90/3
M5
W/L=45/3
W/L=45/3
VSS
Figure.3.16: CMOS operational amplifier circuit with defect 1 introduced using a FIT.
93
VDD
M8 M7
W/L=45/3 W/L=90/3
W/L=90/3
M6
+ Vin
W/L=90/3 W/L=90/3
CC =30FF
M1 M2 VOUT
VE2
M9 M3 M4
W/L=9.6/3 W/L=4.5/1.6 W/L=90/3
M5
W/L=45/3
W/L=45/3
VSS
Fault-Injection Transistor (S-D)
Figure 3.17: CMOS unity gain buffer circuit for S/H with defect 2 introduced
using a FIT.
94
Output of Multiplexer
VE3
VREF
Control Signal
Figure 3.18: CMOS MUX circuit with defect 3 introduced using a FIT.
95
Output of Multiplexer
W/L= 4.5/1.6
VE4
VREF
Control Signal
Figure 3.19: CMOS MUX circuit with defect 4 introduced using a FIT.
96
Chapter 4
(MicroSim Pspice A/D Simulator, V.8) simulations on IDDQ testing of a 10-bit charge
scaling DAC. SPICE level 3 MOS model parameters were used in simulation [42], which
are summarized in Appendix A. The chip was designed using L-EDIT, V.8.03 in standard
1.5µm n-well CMOS technology. The chip occupies an area of 692 µm × 516 µm and
includes 240 µm × 40 µm area of BICS. DAC design was put in 2.25 mm × 2.25mm
size, 40-pin pad frame for fabrication and testing. In the following sections, theoretical
results (simulated from PSPICE) and experimentally measured values will be presented
and discussed. HP 1660CS logic analyzer was used for testing the packaged device
described in Appendix B.
Figure 4.1 shows the layout of a 10-bit charge scaling digital-to-analog converter
with four fault injection transistors distributed across the chip. FIT-1 is injected in the
operational amplifier part of the chip. FIT-2 is injected in S/H circuit of the chip. FIT-3
and FIT-4 are injected in the multiplexer parts of the chip. Figure 4.2 shows the chip
layout of 10-bit charge scaling DAC including BICS within a pad frame of 2.25 mm ×
2.25 mm size. . The area of the DAC alone is 692 × 502 µm2. The entire area of the CUT
along with BICS is 692 × 516 µm2. Therefore the BICS occupies only 242 × 40 µm2 of
the entire chip area. Figure 4.3 shows the microchip photograph of 10-bit charge scaling
97
Defect 4 (FIT-4)
Defect 3 (FIT-3)
Defect 1 (FIT-1)
Defect 2 (FIT-2)
Figure 4.1: CMOS chip layout of a 10-bit charge scaling DAC with four fault injection
transistors distributed across the chip.
98
Figure 4.2: CMOS chip layout of a 10-bit charge scaling DAC including
BICS within a padframe of 2.25mm × 2.25mm size.
99
DAC and BICS for IDDQ testing. Figure 4.4 shows the simulated and measured
characteristics of the 10-bit charge scaling DAC when the faults are not activated. The
equivalent analog output voltage is shown for all the 1024 input combinations. Figure 4.5
shows the measured DNL characteristics of the 10-bit charge scaling DAC when the
faults are not activated. DNL is within ± 0.6 LSB. Figure 4.6 shows the measured INL
characteristics of the 10-bit charge scaling DAC when the faults are not activated. INL is
Figure 4.7 shows the simulated output of the opamp when the fault (VE1)
is activated (Fig 3.16). When the opamp is given a sine wave of 1mV p-p, the output
obtained is a sine wave of 105mV p-p with a gain of 105. Offset voltage is increased to
1.9V from 33µV without faults. Figure 4.8 shows the transfer function with the fault
activated with significant non-linearity introduced in the narrow transition region. Figure
4.9 shows the gain versus frequency response of opamp with fault activated. The
amplifier 3dB gain with the fault activated is 38dB as compared to 77dB when the fault is
deactivated. The bandwidth is increased to 3 MHz from 100 KHz without fault (Fig
2.16).
Figure 4.10 shows the simulated output response of the multiplexer circuit when
FIT-3 is deactivated (Fig. 3.18). When the VCONTROL goes HIGH the multiplexer chooses
the VREF and when VCONTROL goes LOW, the multiplexer chooses GND input. The output
waveform obtained is a pulse of 2.5V p-p. Figure 4.11 shows the simulated output
response of the multiplexer circuit when FIT-3 is activated (Fig 3.18). The output
waveform obtained is 1.7V p-p. Figure 4.12 shows the simulated BICS output when the
100
Figure 4.3: Microchip photograph of 10-bit charge scaling DAC and BICS for IDDQ
testing.
101
DAC output characteristics Simulated
Experimental
2.5
Analog Output
2
Voltage
1.5
1
0.5
0
0 256 512 768 1024
Digital Input Code
Figure 4.4: simulated and measured characteristics of a 10-bit charge scaling DAC.
Note: Faults are not activated.
102
0.8
0.6
0.4
0.2
LSB's
0
0 200 400 600 800 1000
-0.2
-0.4
-0.6
-0.8
Digital Input Code
103
0.9
0.8
0.7
0.6
0.5
LSB's
0.4
0.3
0.2
0.1
0
0 200 400 600 800 1000
Input Code
104
2.5 Gain = 105 Input
Output
2
1.5
Voltage (V)
1 Offset = 1.9V
0.5
0
0 100 200 300 400 500
-0.5
-1
Time (us)
105
3
Output
2.5
2
1.5
1
0.5
0
-0.015 -0.01 -0.005 -0.5 0 0.005 0.01 0.015
-1 Input
-1.5
-2
-2.5
106
Gain = 41dB
3dB Gain = 41dB - 3dB = 38dB
45
40
35
30
Gain (dB)
25
20 3dB Band width = 3MHz
15
10
5
0
1 10 100 1000 10000 10000 1E+06 1E+07 1E+08 1E+09
Frequency0(Hz)
Figure 4.9: Gain versus frequency response of the CMOS opamp circuit with fault
introduced.
107
VREF
2.0V
GND 0V
t
2.5V
VCONTROL
-2.5V
t
Error-Signal
0V
t
2.5V
MUX
Output
-2.5V
t
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
Time
Figure 4.10: Simulated output response of the multiplexer circuit of Fig 3.18
without defect.
108
VREF 2.0V
GND 0V
2.5V
VCONTROL
-2.5V
0V
Error-Signal
2.5v
2.0V
MUX
Output
0.3V
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
Time
Figure 4.11: Simulated output response of the multiplexer circuit of Fig. 3.18 with fault
activated.
109
+2.5V
0V
-2.5V
VENABLE
+2.5V
0V
-2.5V
Error-Signal (VE1)
+2.5V
0V
-2.5V
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
BICS Output
Time
Figure 4.12: Simulated BICS output of the circuit of Fig. 3.16 when Error-signal-1 for
defect-1 is activated.
110
FIT-1 is activated. FIT-1 is a defect injected between the source and drain of the
operational amplifier circuit (Fig.3.16) of the chip (Fig 4.2). In Fig. 4.12, when the
VENABLE signal is ‘high’, the BICS is by-passed and the defect is not detected. When the
VENABLE signal is ‘low’, the BICS is enabled and if the Error-signal is ‘high’, the BICS
detects the faults and the output of the BICS is ‘high’. Figure 4.13 shows the simulated
BICS output when the FIT-2 is activated (Fig.3.17). FIT-2 is a defect injected between
the gate and drain of the unity gain buffer in the S/H circuit (Fig. 3.17) of the chip (Fig.
4.2). In Fig. 4.13, when the VENABLE signal is ‘high’ the BICS is by-passed and the defect
is not detected. When the VENABLE signal is ‘low’ the BICS is enabled and if the Error-
signal is ‘high’ the BICS detects the faults and the output of the BICS is ‘high’. Figure
4.14 shows the simulated BICS output when the FIT-3 is activated. FIT-3 is a defect
injected between the gate and source of the multiplexer circuit of Fig. 3.18 of the chip
(Fig. 4.2). In Fig. 4.14 when the VENABLE signal is ‘high’, the BICS is by-passed and the
defect is not detected. When the VENABLE signal is ‘low’, the BICS is enabled and if the
Error-signal is ‘high’, the BICS detects the faults and the output of the BICS is ‘high’.
Figure 4.15 shows the simulated BICS output when the FIT-4 is activated. FIT-4 is a gate
oxide short injected between the gate and substrate of the multiplexer circuit of Fig 3.19
of the chip (Fig 4.2). When the VENABLE signal is ‘high’, the BICS is by-passed and the
defect is not detected. When the VENABLE signal is ‘low’, the BICS is enabled and if the
Error-signal is ‘high’ the BICS detects the faults and the output of the BICS is ‘high’.
Figure. 4.16 shows the CMOS circuit diagram of a 10-bit charge scaling
DAC with four fault injection transistors. Figure 4.17 shows the simulated output of the
111
4.0V
0V
-4.0V
VENABLE
4.0V
0V
-4.0V
Error-signal (VE2 )
4.0V
0V
-4.0V
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
BICS Output
Time
Figure 4.13: Simulated BICS output of the circuit of Fig. 3.17 when Error-signal-2 for
defect-2 is activated.
112
4.0V
0V
-4.0V
VENABLE
4.0V
0V
- 4.0V
Error-Signal (VE3 )
4.0V
0V
-4.0V
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
BICS Output
Time
Figure 4.14: Simulated BICS output of the circuit of Fig. 3.18 when Error-signal-3 for
defect-3 is activated.
113
4.0V
0V
-4.0V
VENABLE
4.0V
0V
-4.0V
Error-Signal (VE4 )
4.0V
0V
-4.0V
0s 5us 10us 15us 20us 25us 30us 35us 40us 45us 50us
BICS Output
Time
Figure 4.15: Simulated BICS output of the circuit of Fig.3.19 when Error-signal-4 for
defect-4 is activated.
114
STO RAGE UNITY-GAIN
TG SWITCH BUFFER
CAPACITO R
A=1
- +
-
Vo
+
CH -
512C 256C 128C 64C 32C 16C 8C 4C 2C C C
VDD VDD
M8 M7 M8 M7
Output of Multiplexer W/L=45/3
W/L=90/3 W/L=45/3
W/L=90/3 W/L=90/3
M6 W/L=90/3
M6
+ Vin
+ Vin
W/L=90/3 W/L=90/3
- Vin
CC =30FF W/L=90/3 W/L=90/3
M1 M2 Vout CC =30FF
W/L=4.5/1.6 W/L=9.6/3 M1 M2
- Vin Vout
M9 M3 M4 VE
VE W/L=90/3
M5 M9 M3 M4
W/L=45/3 W/L=9.6/3 W/L=4.5/1.6 W/L=90/3
VREF W/L=45/3 M5
W/L=45/3
Control Signal VSS W/L=45/3
VSS
Fault-Injection Transistor (S-D)
Output of Multiplexer
Output of Multiplexer
W/L=4.5/1.6
VE
VE
VREF
Control Signal
VREF
Fault Injection Transistor ( G-S) Control Signal
Figure 4.16: CMOS circuit diagram of 10-bit charge scaling DAC with four fault
injection transistors distributed across different parts of the circuit.
115
2.5V
-2.5V
t
VENABLE
2.5V
-2.5V
t
ERROR-SIGNAL-1
2.5V
-2.5V
t
ERROR-SIGNAL-2
2.5V
-2.5V
t
ERROR-SIGNAL-3
2.5V
-2.5V
t
ERROR-SIGNAL-4
2.5V
-2.5V
t
10US 20US 30US 40US 50US 60US 70US 80US 90US 100US
BICS OUTPUT
Time
Figure 4.17: Simulated BICS output with defects induced using fault injection
transistors.
116
BICS when four FIT’s are activated. In Fig. 4.17, when the VENABLE is ‘high’ the BICS is
disabled and when it is low the BICS is enabled. When the BICS is enabled and if there is
any Error-Signal going ‘high’ the BICS detects the fault. Table-1 shows the faulty IDDQ
values obtained for the different kinds of faults are injected in the chip. Figure 4.18
shows the wave forms obtained form the logic Analyzer. The error signal is kept ‘high’
and the BICS is tested for its operation in the normal mode and test mode. When the
VENABLE is ‘low’, the BICS is in the test mode and if the error signal is ‘high’,
PASS/FAIL gives logic ‘1’ and thus, it detects the fault induced. When the VENABLE is
‘low’, the BICS is in the normal mode, PASS/FAIL gives logic ‘0’ thus, no fault is
detected.
corresponding to four different faults distributed across the chip. The reference current, is
IREF is 1mA. Table 4.1 shows that in the presence of faults IDDQ reaches to 3mA. The
calculated and measured IDDQ values are in close agreement. Table 4.2 summarizes a
comparative study of design of previously reported BICS with the present BICS design.
Our BICS design uses only seven transistors and an inverter buffer. The BICS requires
neither an external voltage source nor a current source. Further more, the BICS does not
require clocks.
117
Figure 4.18: HP 1660CS logic analyzer test results on a fabricated CMOS 10-bit charge
scaling DAC showing the performance of BICS in normal and test modes.
118
Table 4.1 Theoretical and measured IDDQ for different fault types
Note: The reference current, IREF =1mA
119
Table 4.2: Comparison of built-in-current sensors
Note: Mode Select: Normal Mode / Test Mode, Control Pins: VENABLE and EXT.
120
Chapter 5
well CMOS technology. The DAC is tested with all 1024 digital input word
combinations. The unit step is about 1.9mV. DAC operates with ±2.5V supply voltages.
The reference voltage used is 2V for a ‘HIGH’ and ground (0V) for a ‘LOW’. The 10-bit
DAC is used as a circuit under test (CUT). The CUT is tested with a novel built-in
current sensor (BICS), which has a very negligible impact on the performance of the
The present BICS works in two-modes: normal mode and the test mode. In the
normal mode, BICS is isolated from the CUT due to which there is no performance
degradation of the circuit under test. In the testing mode, BICS detects the abnormal
external voltage source nor a current source. The present BICS is designed with only
seven transistors and an inverter buffer. The present BICS design which is based on
current comparison, combined with the novel fault-injection technique, can be used for
IDDQ testing of other type of data converters of the type shown in Fig. 5.1 and is
suggested for future work. The built-in current sensor of the present work requires less
area and is more efficient than the conventional current sensors. It is shown that with the
use of a novel fault injection technique, combined with a built-in current sensor design,
121
O vers am p ling C lo c k
(a)
1-Bit D A C
(b)
122
Bibliography
[1] R. Rajsuman, Iddq Testing for CMOS VLSI, Artech House, 1995.
[2] C.Y. Chao, H. J. Lin and L. Milor, “Optimal testing of VLSI analog circuits,”
IEEE. Tran. on Computer-Aided Design of Integrated Circuits and Systems,
vol.16, Jan. 1997. pp. 58-77.
[3] C.Y. Pan and K.T. Cheng, “Pseudorandom testing for mixed-signal circuits,”
IEEE Trans. on Computer-Aided Design of Integrated Circuits and Systems,
vol 16, Oct.1997. pp. 1173-1185.
[6] B. Dufort and G. W Roberts, “On-chip analog signal generation for mixed-
signal built in self-test,” IEEE J. Solid-State Circuits, vol. 34, March 1999. pp.
318-330.
[8] B.D. Brown. R.D. McLeod, “Built-in current mode circuits for Iddq
monitoring,” Proc. 1993 IEEE Custom Integrated Circuit Conference, 1993,
pp. 30.6.1-4.
[9] C.W. Hsue and C. J. Lin, “Built-in current sensor for Iddq test in CMOS,”
Proc. Int. Test Conf., 1993, pp. 635-641.
[11] J.B. Kim, S. J. Hong, and J. Kim, “Design of built-in current sensor for IDDQ
testing,” IEEE J. Solid-State Circuits, vol. 33, Aug. 1998, pp. 1266-1272.
[12] T.Y. Chang, C. C. Wang, and J. B. Hsu, “Two schemes for detecting CMOS
analog faults,” IEEE J. Solid-State Circuits, vol. 27, Feb. 1992. pp. 229-233.
123
[13] W. Maly and M. Patyra, “Built-in current testing,” IEEE J. Solid-State
Circuits, vol. 27, May. 1992, pp. 425-428.
[14] M. Favalli, P. Olivo, M. Damiani, and B. Ricco, “Novel design for testability
schemes for CMOS IC’s,” IEEE J. Solid-State Circuits, vol. 25, Oct. 1990, pp.
1239-1246.
[15] Y. Miura and K. Kinoshita, “Circuit design for built-in current testing,” Proc.
Int. Test. Conf., Nov. 1992, pp. 873-881.
[16] T.L. Shen, J. C. Daly, and J. C. Lo, “A 2-ns detecting time, 2-µm CMOS
built-in current sensing circuit,” IEEE J. Solid-State Circuits, vol. 28, Jan.
1993. pp. 72-77.
[17] P. Nigh and W. Maly, “A self-testing ALU using built-in current sensing,”
Proc. 1989 IEEE Custom Integrated Circuit Conference, 1989, pp. 22.1.1-
22.1.4.
[18] W. Maly and P. Nigh, “Built-in current testing-feasibility study,” in Proc. Int.
Conf. Computer. –Aided Design, Nov.1988, pp. 340-343.
[20] S.D. McEuen “IDDQ benefits”, Proc. of 1991 IEEE VLSI Test Symposium,
paper 14.1, 1991, pp. 285-290.
[21] J. Argüelles, M. Martínez and S. Bracho, “Dynamic IDD test circuit for
mixed signal ICs,” Electronic Letters, vol. 30, no. 6, 17th March 1994, pp.
485-486.
[24] J. Segura, I.D. Paul, M. Roca, E. Isern, C.J. Hawkins, “Experimental analysis
of transient current testing based on charge observation,” Electronics Letters,
vol. 35, no. 6, 18th March 1999, pp. 441-447.
124
[26] V. Stopjaková, H. Manhaeve and M. Sidiropulos, “On-chip transient current
monitor for testing of low-voltage CMOS IC,” Proc. Design, Automation and
Test conference and exhibition in Europe, March 1999, pp. 538-542.
[27] R.J. Baker, H.W Li, D.E. Boyce, CMOS Circuit Design Layout and
Simulation, IEEE Press 1998.
[28] R.L. Geiger, P.E. Allen, N.R. Strader VLSI Design Techniques for Analog and
Digital Circuits, McGraw-Hill, 1990.
[29] P.E. Allen and D.R Holberg, CMOS Analog Circuit Design, Second Edition,
Oxford University Press, 2002.
[33] J.M. Soden, C.F. Hawkins, R.K. Gulati and W. Mao “IDDQ testing: a review,”
Journal of Electronic Testing: Theory And Applications, vol.3, 1992, pp. 291-
303.
[34] P. Nigh, W. Maly, “Test generation for current testing,” IEEE Design and
Test of Computers, Feb. 1990, pp. 26-38.
[37] J.A. Segura, V.H. Champac, R.R. Montanes, J. Figueras and J.A. Rubio,
“Quiescent current analysis and experimentation of defective CMOS circuits,”
J. of Electronic Testing: Theory and Applications, Vol.3, 1992, pp. 337-346.
[38] K.J. Lee and J.J. Tang, “A built-in current sensor based on current-mode
design,” IEEE Transactions on Circuits and Systems-II Analog and Digital
Signal Processing, Vol. 45, No. 1, Jan.1998, pp. 133-137.
[39] J. Rius and J. Figueras, “Proportional BIC sensor for current testing,” J.
Electronic Testing: Theory and Applications, Vol.3, 1992, pp. 387-396.
125
[40] T.L. Shen, J. C. Daly, and J. C. Lo, “On Chip current sensing circuit for
CMOS VLSI,” Proc. IEEE VLSI Test Symposium, paper 16.2, 1992, pp. 309-
314.
[41] C. Hwang, M. Ismail, J.E. DeGroat, “On chip IDDQ testability schemes for
detecting multiple faults in CMOS IC’s,” IEEE J. of Solid State Circuits, vol.
31, No.5, May 1996, pp. 732-739.
[42] www.mosis.edu.
126
Appendix A
127
Mosis Fabricated Chip Model Parameters (T1AZ)
128
Appendix B
Chip Testability
Figure B.1 shows the 10-bit DAC with BICS in 2.25 mm × 2.25 mm padframe.
Figure B.2 shows the Microchip photograph of 10-bit DAC with BICS. The design
13 Input
14 Output
DC test was performed on the independent inverter module to test if the chip did
not have fabrication problems. Logic ‘0’ is applied at the input pin #13 and output (logic
‘1’) is observed on pin #14. A logic ‘1’ is applied at the input pin #13 and output (logic
8 Positive OPAMP
6 OPAMP Output
9 Negative OPAMP
The opamp is tested by giving a sine wave with 4V p-p to the positive input pin
#8 and the output is observed at the output pin #6. The negative input pin is grounded.
129
B.3 Sample and hold Circuit and Testability
2 Control voltage
4 S/H Input
The sample and hold circuit is tested by giving 4V p-p sine wave. The control
voltage of the sampling capacitor is a pulse of ±2.5V. When the pulse is HIGH, the input
is sampled and when the pulse is LOW, the circuit is in the hold mode. The output is
Table B.1gives the pin numbers and their description to test 10-bit DAC
1 S/H Output
2 Control Voltage
3 VSS
4 S/H Input
6 OPAMP Output
7 VDD
130
9 Negative Input (OPAMP)
10 VSS
24 Error-Signal-2
26 Error-Signal-1
27 VCONTROL
28 VENABLE
29 Error-Signal-3
30 VDD
31 BICS Output
131
32 EXT
33 Error-Signal-4
36 DAC output
37 VSS
38 VGND
39 VREF
40 VDD
1. Supply voltages of ±2.5V is given to the power supply pin numbers of the chip
2. The VENABLE pin (#28) is given a ‘high’ voltage (+2.5V), which makes the BICS
3. The EXT pin (#32) is connected to the VSS (-2.5V) when the BICS is in the
5. The DAC is tested with giving various combinations of inputs to the digital input
132
B.6 IDDQ Testing of the 10-bit DAC in Test Mode
1. The VENABLE is given a LOW voltage (-2.5V) to make BICS operate in the test
mode.
4. When the error signals are activated, faults are injected into the chip and the
5. The PASS/FAIL output is observed. The output of the BICS shows a HIGH value
(PASS/FAIL = ‘1’ = +2.5V) when the faults are injected into the chip and when
the BICS is in test mode. When the BICS is in the normal mode the output is
1. The VENABLE is given a pulse input. When the pulse input is ‘high’ the BICS is in
the normal mode and when the pulse input is low the BICS is in the test mode.
3. The PASS/FAIL output is observed. The output of the BICS shows a HIGH
value (PASS/FAIL = ‘1’ = +2.5V) when the VENABLE is low and the Error Signal
is high. The output of the BICS shows a low value (PASS/FAIL = ‘0’ = -2.5V)
when the VENABLE is high and the Error Signal is ‘high’ since the BICS is
disabled.
133
Figure: B.1 10-bit DAC with BICS in the 2.25 mm × 2.25 mm padframe.
134
Figure B.2: Microchip photograph of 10-bit charge scaling DAC and BICS for IDDQ
testing.
135
Vita
Srinivas Rao Aluri was born on June 20, 1978, in Hyderabad, India. He received his
Technology Inc., Allen, Texas. His research interests include IDDQ Testing of Mixed
136
ISCAS 2000 - IEEE International Symposium on Circuits and Systems, May 28-31, 2000, Geneva, Switzerland
0-7803-5482-6/99/$10.0002000 IEEE
IV- 105
with the digital input word 01100 and 01101 are strongly converters (INL<1/2 LSB) to the total number of try-outs.
correlated since they are both implemented using the same In this way the relation between the unit current standard
current sources. The formula given in eq.1 gives us a rather deviation and the INL-yield is determined (fig.1).
pessimistic view on the INL-yield (fig.1) leading to an However, to obtain the results depicted in fig.1 a large
oversizing of the current source transistors [2] (eq.2) and amount of CPU time is necessary. Running a Monte Carlo
hence of the total chip area. simulation for a high resolution D/A converter takes
several hours and that is a major drawback for this
approach.
IV- 106
P(INL-error) = can be deduced from eq.9. At this point, the INL yield is
no longer described as a sum of probabilities (eq.5) but as
P((Y(1)l Z 0.5 & IY (2)(< 0.5 &
I -I
...& Y ( 2 N 1) < 0.5) +
the possibility that a sample from a normal distribution is
P(IY(2)I 2 0.5 & ...& Y ( 2 N -1) < 0.5) + ...+ (5)
I I smaller than half an LSB. This requirement can be written
as :
P ( k ( 2 N -l)l> 0.5)
IV-107
100
simulated. In almost all cases this procedure gives accurate
results.
Constructing fig.5 using the new formula takes only a few
minutes. The time to write the short program For
MATLAB is so to speak the most time consuming. It is
also worth noting that the time necessary to calculate .the
yield is independent of the resolution of the D/A converter
while the time consumption of the Monte Carlo
simulations “explodes” with an increasing D/A converte:r’s
accuracy.
Furthermore, one can easily conclude from this figure that 4. CONCLUSION
for the design of a high accuracy current-steering CMOS
D/A converter the matching parameters play a significant Since high resolution current-steering D/A converters are
role. A small deviation of the required sigma(I)/I can lead strongly dependent on the matching characteristics of the
to a severe yield degradation. technology in which they are processed, it is important to
know the number of functional chips in a set of fabricated
01
devices. It is proven in this paper that time consuming
Monte Carlo simulations are no longer necessary to obtain
results for the Im-yield with a good accuracy. A new
formula has been presented that directly gives you the
0.01
INI-yield of a current-steering D/A converter in function
of the transistor mismatch parameters of the current
sources without any loss of design time.
______ _ -
0.001 4
I------ I
5. REFERENCES
6 8 10 12 14
number of bits [ I ] K. Lakshimikumar and al., “Characterization and Modeling
of Mismatch in MOS Transistors for Precision Analog
Fig.5 : The unit current relative standard deviation in Design”, IEEE Journal of Solid State Circuits, v01.21, Dec
function of the resolution of the D/A converter for a 30 1986, pp. 1057-1066
yield of 99.7% (+), a yield of 50% ( 0 )and of 10% (m) [2] M. Pelgrom, A. Duinmaijer and A. Welbers, “Matching
Properties of MOS Transistors”, IEEE Journal of Solid
The time to create a figure like fig.5 using Monte Carlo State Circuits, vol. SC-24, Oct. 1989, pp. 1433-1439
simulations in MATLAB is given in table 1. In this table [3] K. Lakshimikumar and al., “Reply to ‘A Comment on :
the results for the INL-yield from 100% to 10% for a Characterization and Modeling of Mismatch in MOS
current-steering D/A converter with different resolutions Transistors for Precision Analog Design”, IEEE Journtzl of
can be found. For all the simulations twenty values for the Solid State Circuits, vo1.23,Feb. 1988, pp. 296
relative unit current standard deviation were taken. This [4] C. Conroy, W. Lane and M. Moran, “A Commenf. on
can be understood as follows. In a first coarse ‘Characterization and Modeling of Mismatch in MOS
Transistors for Precision Analog Design,’” IEEE Journd of
approximation a simulation using 10 values for sigma(I)/I -
Solid State Circuits, ~01.23,Feb. 1988, pp. 294-296
that span a wide range- is run. From the obtained result the [5] J. Bastos and al., ‘‘ A 12 bit Intrinsic Accuracy High Speed
interval for the sigma(I)/I that obtain a high INL-yield can CMOS DAC,” IEEE Journal of Solid Slate Circuits, vo1.33,
be specified. In this interval another 10 points are No.12, Dec. 1998,pp. 1959-1969
IV-108
Characterization of a CMOS Current-Steering DAC using State-Space Models
K. Ola Andersson and J. Jacob Wikner
Dept. of E.E., Linkoping University
SE-581 83 Linkoping, Sweden
MERC, Ericsson Microelectronics AB
Box 1885, SE-581 17 Linkoping, Sweden
-
Abstract Performance limitations on current-steeringdig-
ital-to-analog converters (DACs) are due to 6nite output
impedances, nonideal switches, parasitic capacitances, match- source transistor
ing, etc. In this work we present a dynamic state-space model -Id
of a 14-bit current-steeringDAC which includes dynamic non-
cascade transistor
idealities. Simulation results are presented and compared to
measurement results. The model can be used for fast perform-
ance estimation of DIA converters. t
@)
lout
I. INTRODUCTION
Fig. 1. (a) Ideal, (b) PMOS, and (c) simplified model of MOS current
For high speed and high resolution communication appli- source.
cations the current-steering digital-to-analog converter
P A C ) is suitable [ 11. In this work we focus on the dynamic sistor increases the output impedance [2], but still the source
properties of the current-steering converter. With increasing will have a finite output impedance (Fig. l(c)). The output
signal frequency, dynamic errors, e.g., nonlinear settling impedance contains a capacitive part [11, which will degrade
errors and glitches, tend to determine the performance of the the performance at higher signal frequencies.
DAC. We need good behavioral-level models of the
dynamic properties in order to understand the DAC design B. Theswitch
criteria.
In Sec. I1 we introduce the different nonideal components Using differential signal paths is an effective way of
used in a current-steering DAC. These components are put rejecting noise and distortion, e.g., substrate noise or chan-
together to a circuit level model of the whole DAC, as pre- nel charge injection, if the two paths are symmetrically
sented in Sec. 111. designed [2]. Therefore, differential current switches are
In Sec. IV we present a mathematical analysis of the cir- commonly used and they are implemented with two or more
cuit-level model. The differential equations of the system MOS transistors in parallel (Fig. 2(a)). In this work the
are given on a state-space form, yielding short simulation switch is represented by the MOS switch-on resistance(Fig.
times. The model helps us understand what can be done to 2@)). There also exist parasitic capacitances in the switch,
minimize the influence of dynamic errors. The model is but these are lumped into the parasitics of the current source
Spice-like, but used in Matlab it has an improved modularity and output wire.
and flexibility.
In Sec. V we discuss the application of the model on a C. The Output Wire
14-bit current-steering DAC. Single-tone and multi-tone
properties are discussed. The output wires of the DAC should have zero imped-
ance to reduce the voltage drops. In reality, the wires contain
resistive as well as capacitive and inductive parts. For high
n. NONIDEAL CURRENT-STEERING accuracy, a transmission line model or an RC-ladder should
DAC CIRCUITELEMENTS be used for the wire [3]. However, we want to achieve short
simulation times, i.e., a small number of circuit nodes.
In this work we focus on the influence of unwanted resis- Therefore, we trade accuracy for a lower complexity model.
tive and capacitive parts (parasitics). The parasitics in the As a simple approximation, we use a resistor and capacitor
current sources have one of the most crucial impacts on per- in parallel, i.e., a simple RC-ladder, as shown in Fig. 3. The
formance. Matching is not considered but can easily be impedance includes the internal wire impedance as well as
introduced in simulations of the model presented by simply the off-chip load.
changing the numerical values of currents and parasitics.
A. The Current Source III. CIRCUIT-LEVEL
MODELING
The ideal current source (Fig. l(a)) should have an infi- In this section we briefly present the ideal current-steer-
nite output impedance. A cascoded PMOS implementation ing DAC, followed by a presentation of the nonideal DAC
of a current source is shown in Fig. l@). The cascode tran- constructed with the components described in Sec. 11.
668
h c . 43rd IEEEmdwest Symp. on Clrcuits and System, Lanslng MI, Aug 8-11,2000
0-7803-6475-9/00/$10.~~~~ 2oM)
Fig. 2. (a) Ideal and (b) nonideal differential switch.
%@ +
Fig. 4. The (a) ideal and (b) nonideal current-steering DAC.
a K + 1, i
-
- R ~1j ,c for j = 1, ..., K
and
a 1,. J. = 0 for all other i, j
We want to write the equations on vector form and we
introduce the voltage vector, V ,defined as and the vector B is given by
669
(4 PSD plot for a single-lone input,fsample = 2 MHz
0
-20.
I
E
g -40-
e -60-
-
0
I.......+....... I .g -80-
a
m
a-100-
@oR
l
-
5
Fig. 5. One output channel during a switching period.
-20
The input signal, u(t) ,is a unit step. Since the final values of
the nodal voltages, V i ( T ) ,will become initial values in the
next switching period, all those are needed as output signals
from the state-space system. We also need the output voltage, I I
VJ?),and it is convenient to define the output, Y , as
Y(t) = V ( t ) = X(t) + v(o-)
. u(t) (8)
Equation (5) together with (8) is a standard state-space formu-
lation of the differential equations describing the system [4].
-8""
0.1 0.2 0.3 0.4 0.5
Normalized Frequency
At a first stage, the issue for the simulator is to determine the
output values at multiples of T. Fig. 6. Single-tone spectra for the DAC model using sampling frequencies
of (a) 2 MHz and (b) 10 MHz.
v. SIMULATION RFSJLTS
TABLE I
SMULATION VALUES FOR SOURCES AND PARASIllCS.
The results from Sec. IV enables us to do simulations in a
numerical computation environment, e.g., Matlab, which has Component LSB value MSB value
been chosen for this work. The advantages of using mathe-
matical over circuit-level simulators are for example the sim- Source output resistance 150 GQ 18.3 MSZ
plicity of creating different input signals to the system and the
powerful signal processing tools. Source output capacitance
Quantitative values on parasitics and other nonidealities Source output current 1.22 p A 1OmA
have been estimated through layout extractions and process I I
parameters of a 0 . 3 5 ~CMOS process. The values have been Switch on-resistance I 50051 I 16 51
used to simulate the behavior of a 14-bit current-steering
DAC. In Table I we show the circuit values used in the simula- I Loadcapacitance I 120 pF ~ - -1
tions. Simulations are compared to measurements of a 14-bit
converter with approximately the same properties as given in
I Wire resistance I 100 SL I
Table I. the nonideal and an ideal spectrum. No visible spurious tones
are appearing, only the quantization noise is present. How-
A. Simulated Signal Spectra ever, when the frequencies are increased by a factor of 5, we
We have carried out single-tone simulations of the 14-bit clearly see how the dynamic nonidealities introduce distortion
DAC at different frequencies. In Fig. 6(a) a half-scale single- of odd order. In the separate channels there also exist harmon-
tone spectrum of the differential output signal is shown, where ics of even order that are cancelled in the shown differential
the sampling frequency is f, = 2 MHz . Fig. 6(b) shows the output spectrum. These even order terms become visible in
the differential spectrum aswell, if there is a slight mismatch
situation for f = 10 MHz . In both cases the ratio between between the two channels, e.g., a difference between the load
the signal and sampling frequencies is prime and resistances. In a real implementation we would expect to have
spurious tones even at relatively low frequencies, due to cur-
f signal'f s 1'9 rent source mismatch and other nonidealities not included in
In the 2 MHz case we basically find no difference between the modeling.
670
B. Single-Tone Linearity
PSD plc4 for a multi-tone input, fsample = 10 MHz
One important single-tone linearity measure is the spuri-
ous-free dynamic range (SFDR), which is defined as the
power-ratio between the signal and the largest spurious tone.
In Fig. 7 we show simulated and measured values of SFDR
as a function of signal frequency. The sampling frequency is
fs = 10 MHz . Simulated and measured results show a close
agreement. However, for increased sampling frequencies the
model show worse performance than measurements, e.g., at
f, = 40MHz the simulated SFDR is approximately 10 dE3
worse than the measured. This implies that a somewhat more
complex model is needed to closer resemble reality at high
clock frequencies. The rough model of the wire probably -1 40;
0.1 0.2 0.3 0.4 ( 5
accounts for a significant part of the discrepancies between NormalizedFrequency
measurements and simulations, and how the wire should be
modeled for enough accuracy without to much increase of Fig. 8. Multi-tone spectrum for the DAC model using
complexity requires further investigation. The output imped- a sampling frequency of 10 MHz.
ances and switch resistances has been modeled as linear com-
ponents, although they are nonlinear. This is another source of VI. CONCLUSIONS
model errors.
C. Multi-Tone Linearity A simple state-space model of a CMOS current-steering
DAC has been presented, and used to characterize the single-
In several communication applications a multi-tone input tone performance of a 14-bit DAC. Comparisons between
signal is commonly used, e.g., discrete multi-tone (DMT). In measurements and simulations verify the model to relatively
these cases, the multi-tone power ratio (MTPR) is an impor- high accuracy, and the model explains some important limita-
tant linearity measure. The MTPR is determined using M tions on DAC performance. The concept of mismatch can eas-
tones with equal amplitudes, a , and equally spaced in the fre- ily be introduced in the model by simply changing the
quency domain. One of the tones is left out, and MTPR is numerical values of currents, impedances etc.
defined as the ratio between the power of the used tones, i.e., In some applications a higher accuracy behavioral-level
a 2 / 2 , and the power of the distortion term appearing at the model can be desired. This can be done by introducing, e.g.,
leftout frequency position. This is illustrated in Fig. 8, where nonlinear parasitics and transmission line model of wires.
However, a major advantage of the model is low simulation
M = 33 and a = 2 ” - ’ / ( 2 M ) to avoid signal clipping. times with relatively high accuracy. A refinement of the model
Besides the in-band distortion determining the MTPR, we as discussed above will probably increase simulation times
can also see that the nonlinearities introduce a significant significantly, perhaps to a level where a circuit-level simula-
amount of distortion outside of the signal band. tor, e.g., Spice or Spectre, inight be preferred.
Multi-tone properties are more complex than single-tone
properties. MTPR is strongly dependent on the peak-to-aver-
age ratio (PAR) of the signal, i.e., a higher PAR yields a lower ACKNOWLEDGMENTS
MTPR [SI.In the simulation presented in Fig. 8 there is a ran-
dom phase in each tone for reduced PAR. The authors would like to thank MSc. Niklas U. Anders-
son, Ericsson Microelectronics, for valuable discussions and
SFDR vs Signal Frequency,10 MHz sampling frequency all help with DAC measurements.
80, I
\
REFERENCES
75 -
M. Gustavsson, J.J. Wikner, and N. Tan, CMOS datu converfers for
- communications, Kluwer Academic Publishers, Boston, U S A , 2000,
e
370
ISBN 0-7923-7780-X
-----
E
“65-
simulated ‘,
D.A.’Johnsand K. Martin, Analog integrated circuit design, John Wiley
and Sons, New York, U.S.A., 1997, ISBN 0-471-14448-7
Z.F. Jin, J.J. hurin, and Y. Savaria, “A new approach to analyze inter-
6o
55‘
1o5
t 1on
I
10’
connect delays in RC wire models,” in Proc. of the 1999 IEEE Int’l
Symp. on Circuits and Systems (ISCAS’99), vol. 6, pp. 246-249,
Orlando, FL,USA, May 30 - June 2,1999
Signal Frequency [Hz] T. Glad and L. Ljung, Reglerteori - jervuriabla och olinjara metoder,
Fig. 7. Simulated and measured SFDR as a function of the signal Studentlitteratur, Lund, Sweden, 1997, ISBN 91-44-00472-9
frequency. The sampling frequency is 10 MHz. AN971 8, Application Note, Intersil Corporation, April, 1997
671
Clock Jitter Compensation for Current Steering DACs
Andreas Wiesbauer, Dietmar Luis Hernandez Daniel Gruber
Straussnigg, Richard Gaggl,
Martin Clara Universidad Carlos III Telematics / Network Engineering
Madrid, Spain Carinthia Tech Institute, CTI
Infineon Technolgies Austria AG luis.hernandezguc3m.es Klagenfurt, Austria
Villach, Austria daniel.grubergedu.fh-kaernten.ac.at
Abstract-Clock jitter is an important source of error in high- digital means. Simulation results and the simulation setup are
speed current-steering D/A converters. A technique to discussed in section IV, revealing the required hardware
compensate these errors is introduced. Simulations show a complexity to significantly reduce clock jitter sensitivity for
significant reduction to clock jitter sensitivity for the example a 30MHz analog bandwidth, 12bit resolution SD DAC. We
of a Sigma Delta DAC with an analog bandwidth (ABW) of 30 consider random timing errors as the dominant timing error
MHz clocked at 360 MHz. throughout this paper, even though the method can handle
nonrandom jitter as well.
I. INTRODUCTION -------------------------------------
' ~~DAC
analogue signal
is well
It known, that continuous-time sigma-delta (CT digital analogue signal (+ errorej) corrected)
SD) ADCs suffer from sensitivity to clock timing errors signal DAC
caused dominantly by injection of these errors into the sigma
delta loop via the current steering DAC in the outermost ------------- -es---------
estimated
estimated
digntal
error esimal
~ 'esfimated
eo
feedback branch [1]. A model and an analytical description Clk
of the jitter error effect were described in [2]. The jitter esfimated
sensitivity is inherent to any current steering DAC either r---- errorsignal
used inside a CT SD ADC or as a stand-alone DAC. i
' j i er
ero
error
measured
clock jitter measured jitter error
(analogue) clock jitter measuring estimation
ADC(digital)
(digital) estimatedII
~~~~~~~~~~~~~~~~circuit
10 X jitter error
Figure 4. Noise-shaper DAC withjitter compensation
differentiated It follows from the linear model of a noise-shaper that an
digital signal error EJ(z) subtracted from the signal in the feedback loop of
digital T W
signal the noise-shaper will contribute to the output Y(z) of the
noise-shaper in the form
Y(Z) = X (Z)+ Eq (Z)(1 - R(Z)) - Ej (Z)R (Z) ' (2)
where R(z) is the transfer function of the loop filter in the
Figure3. Clockjittererrorestimation feedback branch as depicted in Fig. 4 and Eq(z) is the
quantization error. If properly dimensioned, (11-R(z)l)
<< 1
B. Jitter Error Estimation in the band of interest resulting in
Jitter error estimation is based on (1) and requires Y(z) X(z) - Ej (z) for f < (fample 1(20SR)) (3)
,
5376
IV. MODELLING AND SIMULATION RESULTS Additionally to the measured values the theoretical SNR
values achievable with the given long-term clock jitter were
The clock jitter sensitivity of a DAC with a noise-shaper calculated using the formula derived in [6]
for the wordlength reduction from 12bit to 6bit was K
evaluated. A clock of 360MHz for the DAC and an SNR=-20 logl 2zf ig " I|. (4)
oversampling ratio of 6 are assumed (30MHz bandwidth). y2fJ
The noise-shaper was designed for a peak Signal-to-Noise
Ratio (SNR) of 72dB. 80
Fig. 5 shows the PSD ofthe output signal ofthe modeled I III I I I
DAC including the jitter error and compensated with the 70 7QT
estimated jitter error. Furthermore the PSD of the modeled H \\
clock source, which resembles a typical PLL-type on-chip 60 - FT - .
-
rT- -1 T1 rFITr-
T F \-IrI T T -TI-T -I-
clock synthesizer [5], can be seen on the top right side. _ -_ _ -_ i < L- 1 -_-
504
0 __ _I_I liii I liii IQI liii I
~~~
~~~l II~>II
24011 signa -o
compensat on
-120
'
, 4
theoretical SNRfor givenGl1t 1 1 1 1 1 11 i1
i
-YF ,1 3 fi l t iM 30 1-iH~
II ~ -I-t -i -i -ig+ i_IIII i_ _4 __ _IIIII
+-- -i t=- =
I11
4 I11 I11
-4 ;~~=~~-=
'£j -812 -1 -^4410 -9Cit S
c20-ompina
9I~~~~Jrnl w/n w/ cmpesain
--------- -------------
-m ~~~~~~~~~~~f1010
-2MAOmignlwith 2orMpendtlinaoihno pnsto 0 e1i=0 10l0.X
-140>
Worcie of ban
o.; 0-0W ntflterst It
B. Quantized measured jitter
0 2 4 6 8 10 12 14 16 The second set of simulations includes a more realistic
lequency [Hz] X1 measurement of the clock jitter values. These were sampled
and quantized, simulating an analogue-digital conversion of
Figure 5. Spectrum ofFthe signal with jitter error and the jittered clock the clock jitter values.
(upper right corner)
pensation
corr
To show the impact on the performance of the w.
proposed
correction method several simulation scenarios were 70r-7alisi--c- K ----- V
investigated. As a size for the jitter intensity the long term me o co j- vals T S
rms-jitter (ola) was used. To measure the performance ofqthe 65 simlain a analogue-digital of
proposed jitter error compensation method the SNR vs. clock - - - -- -- - -
60t0 -F-Aes.- - + i \ <K- K
jitter is evaluated. corner)
- 1 4 1-G1 1 1 111 I
A. Idea/measurement and compensation z 50 - - -I -I-~-~~6
-~ -I - - A- - - 4 - +A- - - ---
For the first set of simulations ideal measured clock jitter I I 1111 I I II IIlll i
lXlI
values for each sample instance were assumed. The result of 45 -h--I_I- e-l - 4
these simuationsl aoncompensationFig.o6
tneseslmulatlonscanseseenmrlg.O. 40 \al jtralgeo.
wloo cl o |
400 clock jittersamplingresolution:2bic -
.. . . . ,. .|
,~~~~~~~~~~~clock
>| jitter sampling resolution: 3 bit i i i i i
At low jitter values quantization noise limits the SNR. 35 - clock jitter sampling resolution: 4 bit 4 L -_ _--
-
With increasing clock jitter, starting at 2ps, the SNR of the-r signal with ideal compensationI I I
uncorrected signal decreases linearly with the logarithm of 10-11 -10
the long-term clock jitter. The signal with compensation 10 t[si 10
shows much better performance up to high clock jitter
intensity. The decrease of the SNR at a long-term clock jitter Figure 7. SNR vs. clock jitter for different ADC resolution
intensity of about l00ps results from limitation of the
compensation values the values in the feedback loop may The modeled jitter error is no longer exact and introduces
not exceed reasonable size. The compensation does no additional quantizationnoiseinthe feedbackloop. Therefore
longer cancel the error fully but does still decrease it. the performance of the compensation deteriorates with
decreasing resolution of the ADC converting the clock jitter
values. Fig. 7 shows the SNR vs. clock jitter for different
5377
C. Influence of measurement errors
1) Gain error
Em
70
65 - - - --T-T
60
55
- -
----
-
I
I
I
l111
I-
.\ ~ ~ ~ ~ ~ ~ I I1 II\ I \
ADC resolution. The ADC had full scale at 5 times the
standard deviation of the cycle-cycle clock jitter values
which covers over 99°0 of the possible clock jitter values.
1111111
IIlulL
III
1111
1
I1-
Iit--
I
I
X1111I
lI
I___
-0
1 11 1 1Il
l11l l
I ~~~~~~~II I1I1I1I
--t
-
g
IN
4-11-
- Ii I II- I
-t
1
l l l l l l l l l
t
_ \_
x
I
--4\_-
_T 7
IC
compensation. The resolution of the jitter measurement is 8
bits lower than the converter resolution. Compared to the
main DAC the measurement DACs and integrator noise and
accuracy specifications are relaxed by a factor of 256,
leading to an almost negligible cost adder.
m C
l
EE
Zl I0-
75
704
65
60 -
35
..=
I
-1-
-
- - -
111
I
- - -i-
1-
---1-1-5IIiIK- I _
I_ ___
1
--
45 - --_ - -___I-
I
I I II-1114
- -I-I-I-I HI
-
I 1111111
--
I
-1-I
I I
__ I _ II Al--
-
I I
1-1-1
10
40
7
1-1-lX1-4
H--
10
4 I
I I ii
[s]
-HI I
-1-I 4-4
I4
-
liL
3
I
I
4X41
_I I I
I I I III
4-4 4~~~~~~~~~
4
1JA\I-1-0
1 1
W-
---]
4-I--- --]
-lo
-I_]
11WWWW
I
X
rr I I I1I111 V. CONCLUSIONS
z 50 - -- 1 11111
1 I 11It-+
+Ht --t
A method for compensating clock jitter errors was
- - + + -
41 I41
b I14
45 ---I--I--I-HI-HI-----H--t-H4-4I
45
I
--I
4I-
-4 A
5378
CLOCK-JITTER INDUCED DISTORTION IN HIGH SPEED CMOS
SWITCHED-CURRENT SEGMENTED DIGITAL-TO-ANALOG CONVERTERS
Jose' Luis Gonzdez and Edtrard Alarcdn
1-5 12
0-7803-6685-9/01/$10.0002001 JEEE
complement the input bits and distribute them to several decoder implemented using a binary array of switched current sources. In our
gates. For example. in the vicinity of the crossover point of a analysis all the non-idealities that may be included in the DAC model.
sinusoidal input, one sample should be 01 11111111111111 and the such as current source finite output resistance, different switching
next should be 1000000000000000.If the thermometer code segment times. parasitic elements and mismatching. are hmed off. since we
has a length in = 8 bits, there will be a total of 8 simultaneous are interested only in the clock-jitter effects. In this case the binary
switching large fan-out inverters at the input of the decoder. output structure can be used for both binary and thermometer
Meanwhile, at the decoder output only one of the 256 bits will tum segments. The DAC incorporates a first order reconstruction output
on. filter with a cut-off frequency equal to half the sampling frequency.
2.2 Clock jitter produced by the dUdt noise. The exact time at The sampling frequency isfs = 105 MHz (sampling period T, = 9.52
which the analog output signal switches is fixed by the clock signal us). The converter is sized to provide a differential output with a 2 V
edges. The clock driver generates the internal clock signal from an maximum amplitude across a 50 R load resistor. The exact sizes of
external clock. The digital input codes are supplied to the the perfectly scaled output current sources are not relevant for the
synchronized with the external clock. The input bits changes generate results of the analysis because they are considered ideal current
Wdt noise at the internal power supply nodes, as explained before. sources without mismatching.
That noise is either coupled directly to the clock driver if it shares the
same digital power supply or it is coupled through the substrate if it
does not. The dVdt and/or substrate noise affects the delay of the
output driver. generating clock-jitter [SI. Figure 2 shows the
simulated delay change of a clock driver due to Wdt noise generated
by the decoder when multiple bits of the input code simultaneously
change. The simulation w a s performed using HSPICE with a 0.5pm
technology transistor models [9] and a typical clock driver circuitry. Fig. 3.DAC siinulation set-up
not shown here for confidentiality reasons. The on-chip decoupling The clock-jitter is introduced in the ADC block that generates the
capacitance and damping resistance are S pF and 1 R. respectively. clock signal for the DAC by adding a variable delay to an external
and the power supply package pins inductance is varied from 3 nH to clock signal (see figure 3). This delay can be of two different types.
10 nH. The triangular current pulse used to model the decoder activity One type causes a random clock-jitter. The other type causes input-
is 500 ps wide and its amplitude is varied from 5 mA to 100 mA. The dependent clock-jitter and it is calculated from the number of TCS
input data is changed 3 ns before the clock rising edge. Even if the bits changing from sample to sample using the model presented in
current pulse due to the input data change is not simultaneous with section 2. The number of TCS bits that change is multiplied by the
the external clock edge (as is the case here). the damped oscillation jitter amplitude. that is a parameter of the simulation.
generated by the dUdt noise pulse in the package-chip RLC resonant In the case of the random jitter a uniformly distributed random
circuit [lo] affects the internal clock edge delay. Different noise number is generated for every sample and the sample's clock delay is
levels from sample to sample produce different delay changes in the calculated from that number. The amplihlde of the distribution of that
clock driver and this is the cause of the internal clock-jitter. This random number is also a parameter of the simulation.
effect is much more pronounced than that due to thermal noise. as In order to determine the effects of clock-jitter in the output signal
shown in [SI. distortion and the relation ~ t theh DAC segmentation struchwe. a set
h the following analysis it will be used a linear expression relating of simulations are performed. The parameters of the simulations are:
the clock-jitter amplitude and the number of input bits simultaneously
switching based on the explanations of subsection 2.1 and the results
input frequency v,,). type of clock-jitter (random or input-dependent).
number of TCS bits ( n i that ranges form 0 to 16) and clock-jitter
of figure 2. Only the changes in the TCS bits are relevant. In this way. amplitude. In the case of the random clock-jitter. the amplitude is
an increase in the size of the thermometer segment I I I will raise the defmed as the 50% value of the positive uniform random distribution.
maximum clock-jitterand produce a lower SFDR. as s h o w below. In the case of the input-dependent clock-jitter the amplihrde is defmed
w as the slope of the linear relation between the number of bits changing
m from sample to sample and the delay added to tlie clock signal (i.e. an
amplitude of 0.5 ps/bit will produce a change in the clock signal delay
of 2.0 ps if 4 of the 111 input bits of the thermometer segment switch
simultaneously).
Three different input frequencies are selected (1.08 MHz. 5.13 MHz
and 10.9 MHz). Transient simulations are perfoimed containing SO
cycles of the input sinusoidal signal. Then the FFT of the analog
outpnt signal is obtained taking 131072 samples and usiug a H&g
window to obtain the frequency spectnun of the output signals and to
Fig. 2. Clock signal delay changes caused by dVdt noise. measure the SFDR.
1-5 13
An increase of a factor of 10 in the jitter amplitude produces a 20 dB
increase in the noise floor. There is also depeudeuce with the input
frequencykampling frequency ratio. For lower ratios. the noise floor
is lower than for higher ratios.
Uniformelydistributedrandom clock jitter 16 bit DAC
f (AUsvAtchingt !s)
-40 1 P 80 I--to.5ps
a
70 -
U)
60-
- m
40 10 15 20
17- ’lo
100 Clock-jitter
amplitude
around 6 to 7 simultaneous switching bits for the highest input For higher 111 some iuput codes are randomly skipped from the
frequency and around 5 to 6 for the lowest freqneucy. This means that ordered sequence (if there is no relatiou between the input and the
sampling frequency) and the number of TCS bits that change from
the saturation point (the m value where the corner iu the SFDR ciirves
sample to sample behave as a random process.
appears) depends on the input frequency.
1-5 14
with the same effects on DAC performance show previously. There
fin= 10.9 MHr fin=1.03MHr
is a linear dependence with jitter amplitude (a 10x factor in jitter
amplihide reduces 20dB the SFDR) and with input frequency
(approximately 3dB of SFDR are lost for every 2x increase in the
input frequency).
The results obtained in this work reveal the importance of limiting the
1000 clock-jitter in high-speed DACs to values below 0.1% of the sampling
50 500 period by avoiding large Wdt or substrate noise that causes clock-
4m0 2 4 6 8 10 12 14 16 3ooo0 2 4 6 8 10 12 14 16 jitter and distorts the output signal. The influence of ni (the number of
350 thermometer code segment bits) is only important for small
300 segmentation ratios (nr 5 4). For larger segmentation ratios. SFDR
250 shows no dependence on nz. Several techniques can be used to reduce
200
150 1000
the dvdt or substrate noise. but they are out of the scope of this paper.
100
500
Interested readers are referred to the bibliography [7]. The
50
0 dependence with the segmentation ratio that has been found in the
4oo0 2 4 6 8 10 12 14 16 3ooo0 2 4 6 8 10 12 14 16 case of input-dependent clock-jitter is proposed as an additional
design criteria that can be used to find the optimal segmentation for
250 2000 high-speed current-steering DACs.
Acknowledgements. Work supported by CICYT Project No. TIC-469.
and by Fulbright fellowship FU 1998-0038446010.The authors would
like to thank David Anderson and Paul Hu, from Motorola, Phoenix (AZ)
and Prof Olgierd A. Palusinski form University of Arizona. Tucson (AZ)
for providing the simulation models of the driver and Guido Stehr and Tao
Wu for helping with the DAC SABER modeling.
7 250
1000
1500 k REFERENCES
?LL&.l
2000
[2] Marcel J.M. Pelgrom, Aad C. J. Duinmaijer, and Anton P.G. Welbers,
“Matching Properties of MOS Transistors”. IEEE J. Solid-Store Circriifs,
%1
.,
0 500
Vol. 24, NO.5, Oct. 1989,pp. 1433-1440.
[3] David A. Johns, Ken Martin. Analog Integoied Circriit Design. John
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Wiley& Sons. Inc. New York, 1997.
Number of swltching bits inside the thermometer code segment 141 Geet A. M. Van der Plas, Jan Vandendbussche,Willy Sansen, Michel S. J.
Fig. 6 Nstograin of the number of swtchmg bits causlng clock-jitter Steyaert, and Geroges G.E. Gielen, “A 14-bit Intrinsic Accuracy Q’
Random Walk CMOS DAC, IEEE J. Solid-Side Circtiit.y, Vol. 34, No.
m the l6-blt DAC for two different mput frequencies and for vanous
12,Dec. 1999,pp. 1708-1717.
thennoineter code segnent lengths [ 5 ] Chi-Hung Lin and Klaas Bult, “A 10-b, 500-Msaples/sCMOS DAC in 0.6
“2..
, IEEE J. Solid-Strite Circiiits, Vol. 33, No. 12, Dec. 1998, pp. 1948-
1958.
5. CONCLUSIONS 161 Yonghua Cong, and Randall L. Geiger, “Switching Sequence
Optimization for Gradient Error Compensationin Thermometer-Decoded
DAC Arrays“, IEEE Tr. on Circriits rind .y\srenn-II, Vol. 47. No. 7, July
hi this paper the effect of clock-jitter in the performance of high- 2000. pp. 585-595.
speed segmented cuurent-steering DACs has been s h o w . The clock- [7] Xavier Aragones, Jose Luis Gondez. and Antonio Rubio,. Ano!isis ond
jitter affects the performance either by increasing the noise floor and Soliitions for Siiitching Noise Coripling in Mixed-Signril ICs. Kluwer
thus reducing the number of significant bits or by reducing the SFDR. Academic Publishers.Boston, 1999.
When a randoin process independent from the input sequence [SI F r d Herzel, and Behzad Razavi, “A Study of Oscillator Jitter Due to
originates the clock-jitter. the noise floor increases linearly with the Supply and Substrate Noise”, IEEE Tr. on Circriits rind Sistenrs-ll, Vol.
46, No. 1, Jan. 1999,pp. 56-62.
jitter amplitude. There is also a dependence with the input frequency: [9] MOTOROLA 0.5pm proprietalyprocess.
for a fixed sampling frequency. higher input frequencies are more [IOIPatrik Larsson, “Resonance and Damping in CMOS Circuits with On-
affected by clock jitter than lower frequencies. Chip Decoupling Capacitance”. IEEE Tr.on Circriits rind Ststenid, Nol.
The clock-jitter is also produced by the switching noise generated in 45, No. 8, Aug. 1998,pp. 849-858.
the binary to thermometer decoder. The noise generated by the [11]Guido Stehr. Behcwiorircil Modelling of o High-precision, High-speed
decoder modifies the delay of the clock driver. The decoder switching Digitcil-to-rincinlog Converter, Master’s Thesis. Universitat Karlsruhe
uoise depends on the input sequence. This. the dVdt originated clock- (TH). Germany, 1999.
jitter depends on the input sequence and on the decoder size. Larger [12]Sober Sinriilrrror Ver 5.0, Analogy Inc.. Beavrton, OR (USA)
[13] Selim Saad Awad, ‘Analysis of Accumulated Timing-Jitter in the Time
segmentation ratios produce larger switching noise because the Domain’, IEEE Tr. on Instrrinientritionrnid Memrirenrent, Vol. 47, No. 1,
decoder comprises more bits. ln this case. dependence between the Feb. 1998,pp. 69-73.
SFDR and the thermometer segment size has been found. An increase [14]Ham0 Kobayashi, Masanao Moriniura, Kensuke Kobayashi. and
in the thermometer code segment bit count ( i n ) reduces the SFDR up Yoshitaka Onaya, ‘Aperture Jitter Effects in Wideband Sampling
to a certain ni value. For higher i n values. SFDR reduction saturates Systems’,Proc. of IMTC‘99 Conference, 1999. pp. 880-885
and does not depend on in. Thus the input dependent clock-jitter for
large segmentation ratios can be considered as random clock-jitter
1-515
Combining DACs for Improved Performance
K. Ola Andersson1,2, Niklas U. Andersson1,2, Mark Vesterbacka1, and J. Jacob Wikner3
1Dept. of E.E., Linköping University, 581 83 Linköping, Sweden, fax +4613139282
2LiDC, Ericsson Microelectronics AB, Box 1554, 581 15 Linköping, Sweden
3SDC, Ericsson Microelectronics, Westmead Dr., Swindon, SN5 7UN, UK, fax +441793490151
E-mail: {olaa, niklasa, markv}@isy.liu.se, jacob.j.wikner@mic.ericsson.se
Summary For higher signal and updating frequencies dynamic er-
This work is an overview of recently proposed methods rors tend to be the limiting factor on the linearity. One
on combining DACs in order to improve performance. source of dynamic nonlinearity is the nonzero output ca-
Some further development of these techniques are also pacitance of the current sources, causing signal depend-
presented. The techniques aim at reducing glitches and ent settling errors, which in turn yields an overall
sensitivity towards limited output impedance in current nonlinear behavior. This phenomenon was modeled with
sources. Matlab in [5]. The equivalent circuit representation of
that Matlab model is shown in Fig. 1(b). It includes the
Keywords: D/A conversion, glitches, nonlinearities
output capacitance and resistance of the current sources,
1. Introduction switch resistance, wire resistance and capacitance. In
Today’s telecommunication applications require high- Sec. 4 we give an example on how to use two sub DACs
speed and high-resolution data converters. In, e.g., digital to reduce the distortion due to limited output impedance
subscriber line (DSL) applications the requirements on by introducing redundancy which allows the common-
the digital-to-analog converter (DAC) is in the order of mode signal to be varied.
12-14 bits of resolution and several MHz of signal band-
2. General multi-DAC concept
width. These requirements are not always easily fulfilled,
Assume that we have an N -bit DAC consisting of a set of
and in this work we give an overview of some recently
M parallel sub DACs, with their outputs summed to pro-
proposed techniques utilizing several sub DACs com-
duce the total output. The number of weights in the sub
bined into one DAC for improved performance.
DACs are chosen so that the overall N -bit resolution can
The current-steering approach (see Fig. 1(a)) is often
used in high-speed applications, since the architecture en-
ables fast operation. The performance is however often I0 Ij IN-1
limited by linearity problems. Mismatch between the
transistors in the current sources causes errors in the bit b0 bj bN-1
weights and limits the static linearity of the DAC. To
solve this problem calibration or randomization tech-
I+ I-
niques, so called dynamic element matching (DEM, see A+ A-
e.g. [1]), are often suggested. In some applications distor-
tion is more crucial than the noise, especially in oversam- RL RL
pled systems where much of the noise typically can be (a)
filtered out. DEM techniques utilize redundant codes and
randomization to transform distortion into noise with the
benefit of linearizing the device but introducing more
switching activity and hence noise. There are schemes to
minimize the glitching activity and still maintain a rea-
sonable amount of randomization [2, 3]. It is also known
that the glitching activity is reduced by using special
codes, such as segmentation of the MSBs or other ar-
rangements. In this work we overview different ways to
combine a pair of converters to reduce the glitches to a
relatively low (digital) hardware cost. In Sec. 3 we high- (b)
light an example where the converters are linearly coded
[4] yielding low glitch errors. Figure 1: (a) Ideal and (b) model of current-steering
DAC including parasitic resistances and capacitances.
be met. The input to each sub DAC is denoted X m . We
Linear 0
DAC 1 A+
let the outputs of the sub converters, A m+ and A m– , be dif- X(n) |X|
encoder
0
0 Ij=(j+1)I0 A
0
ferential, where A m– = A mmax – A m+ . A mmax is the maxi- –
150
area requirements and power consumption
lin
100 A+
X1(n)
DAC1
50 A-
X(n) r(n)
0 Vdump
-8000 -6000 -4000 -2000 0 2000 4000 6000 8000
X
A+
X2(n) I2 I1
Figure 3: Glitching in 14-bit DACs with (a) DAC2
conventional 7-bit segmented architecture and (b) the A- RL RL
proposed linear-coded architecture.
4.1 Architecture overview
The single-ended output currents for the ideal current- Figure 4: Proposed redundant differential DAC
steering DAC in Fig. 1(a) are given by architecture.
I+ = Iu ⋅ ∑ b i ⋅ w i and I – = I u ⋅ ∑
N–1 N–1
b i ⋅ w i , (1) 4.2 Simulations
i=0 i=0 One choice of r(n) discussed in [6] is a random dither
and the differential output current by signal with a given amplitude. This choice of r(n) yields
N–1 N–1 a form of DEM. Another choice that was mentioned in
I diff = I + – I – = 2 ⋅ I u ⋅ ∑ bi ⋅ wi – I u ⋅ ∑ wi (2) [6], is to choose r(n) as the smallest constant signal
avoiding overflow of X 1 and X 2 . In this way the parasit-
i=0 i=0
ic load at the output nodes is lowered. This approach is
where I u is the unit current, b i is bit i of the input code
only applicable to a limited class of signals, i.e., the ones
X , b i is the inverse of b i , and w i is the weight of bit i .
that do not utilize the possible input range, and need to es-
The proposed redundant architecture is shown in Fig. 4, timate the maximum value of all future samples of the in-
composed of two parallel, nominally identical, current- put signal. In this work we modify this approach and
steering DACs (DAC1 and DAC2). The input signal to instead choose r(n) as the smallest integer fulfilling the
DACi is denoted X i(n) , where boundary conditions
X 1(n) = X (n) + r(n) and (3) r(n) ≥ – min(X (n), X max – X (n)) = r min(n) and (5)
X 2(n) = X max – X (n) + r(n) . (4) r(n) – r(n – 1) ≤ ∆r max (6)
Boundary condition (5) ensures that overflow is avoided, 3−tone PSD for conventional DAC
and ∆r max is chosen such that the common mode varia- (a) 0
tions can be sufficiently rejected in following circuitry (in
the previous approach, ∆r max = 0 ). The boundary con- −20
ditions should hold for all n , so we need to keep track of
a few future samples of X , otherwise we risk choosing −40
PSD [dB/Hz]
r(n) too small to fulfil both (5) and (6) for some future
sample. However, since we need to keep track of less −60
samples than before, this approach is more suitable for an
actual implementation. −80
Parameter Value
−60
Output resistance (unit current source) 1 GΩ
Output capacitance (unit current source) −80
5 fF
Switch resistance (all switches) 100 Ω −100
Biasing
A current mode track-and-hold circuit designed for IB,LSB IB,MSB
1 : 256
deglitching a 14-bit current steering DAC is presented.
The deglitcher is based on a developed highly linear cur- LSB MSB
rent memory. The prototype circuit is designed using a 8 Binary Weighted 2 Binary Weighted
+
Traditionally the applications of high-speed DACs Figure 1. Block diagram of the DAC.
have been in video and computer graphics applications,
but recently the migration to wideband wired and wire-
less telecommunication standards and the evolution of ra- optimal control waveforms for the switches, and the use
dio transmitter architectures toward the software-defined of return-to-zero-type output to suppress the output dur-
radio have created a need for high-speed, high-resolution ing the code changes. The return-to-zero technique uti-
telecommunication DACs. lized in [2] yields a clear improvement in high-frequency
In the past the research and development of DACs SFDR compared to earlier reported DACs, but still has
have been heavily concentrated on improving the static some limitations, such as the difficulty of providing large
(DNL, INL) and, to some extent, the time domain specifi- amplitudes to a low-resistance load, the complicated cir-
cations (settling time, glitch area), almost totally neglect- cuitry needed to handle signal dependent parasitics, and
ing spectral purity and other frequency domain character- sensitivity to clock jitter, which is not relaxed, unlike in
istics that are essential in telecommunication devices. conventional DACs, when signal frequency is decreased.
Practically all high-speed DACs are based on the cur- To alleviate the first two of these problems the same
rent steering architecture, one of the main reasons for authors have proposed the track/attenuate technique [3],
this popularity being its capability of driving resistive which is basically a switch put in parallel with the load to
loads without buffering. A 14-bit static linearity has short the output during the DAC switching.
been achieved by using trimming, self-calibration [3], To avoid the jitter problem and signal attenuation it is
and even intrinsically [1]. A typical problem in these possible to use a track-and-hold circuit as a deglitcher;
DACs is the rapid increase in harmonic distortion when the DAC is cascaded with a T/H circuit which tracks the
the signal frequency is increased. This is mainly due to DAC output when it is in steady state and holds a sampled
the glitches occurring at the code changes. The glitches voltage during DAC settling. Although the deglitcher
are results of incoherent timing of the current switches, does a good job of removing code-dependent glitches it
non-optimal shape of the switch control waveforms, and typically cannot achieve as high a speed as a current-
coupling of the digital signals to the analog output. steering DAC and, furthermore, the voltage output pro-
Attempts to reduce glitches include the use of latches vided by the T/H needs to be buffered in order to drive
to synchronize the switch controls, circuits to generate resistive loads.
1 This work was supported by Nokia Networks and the Finnish Na- The deglitcher presented here is based on current
tional Technology Agency (TEKES). mode circuitry. The output is provided in the form of
479
a current which eliminates the need for a buffer. A high Digital input IOUT+
2. Architecture
Clk-
Clk+ Clk+
The core DAC, shown in Figure 1, is based on seg-
mented architecture. The current sources are divided I3 I4
into two unit current source arrays, with 8 LSBs in one Clk1 Current
memory
Clk2 Current
memory
480
output. In addition, a cascode current source is added in triode, not the saturation, region. There, the drain current
parallel with the DAC to bias the current memory. The is given by
base currents of the switch transistors—as long as the
Cox W VDS
2
transistors in the switch pair are matched—are not a prob- ID = (VGS VT ) VDS : (1)
lem, since the switches are enclosed in the feedback loop L 2
when the current is sampled. Now, if the drain-source voltage is kept constant, the cir-
cuit is perfectly linear. To make the voltage on the drain
3.1. Current Memory as constant as possible M1 is cascoded with the bipolar
Typically, the current memories are based on a MOS transistor Q1, which has an inherently large gm, which
transistor whose gate capacitance is used as an internal is further boosted by using regulation. To achieve a high
storage element, while the memory input and output are linearity it is necessary to bias the transistor M1 deep in
both the drain current. the linear region. The cascode current source, consisting
As regards high resolution applications, the most se- of M2 and M3, is for biasing M1.
vere limitation of this circuit is the harmonic distortion The sampling speed of the deglitcher is determined
originating from the sampling switch charge injection. by the time constants associated with the feedback loop.
The nonlinear relationship between the gate voltage and When the loop is broken at the gate of M1 (leaving the
the drain current results in a situation where even a con- capacitor C1 on the output side) there is only one high
stant charge injection produces harmonic distortion in the impedance node (the DAC output), which justifies the
output current. Moreover, the switch operates against use of the single pole approximation, giving the follow-
the gate voltage (V G ), which makes the charge injec- ing gain-bandwidth product:
tion signal-dependent. Ways of reducing the effect of gm1
GBW = ; (2)
the charge injection range from the use of differential cir- C1 + CDAC
cuitry and dummy switches to the very accurate S 2 I tech- where gm1 is the transconductance of M1 and C DAC the
nique [4]. DAC output capacitance. Since the transconductance (to-
In this design the approach taken to achieve 14-bit gether with the full-scale output current) determines the
resolution is two-fold. First, the linearity of the current voltage swing on the memory capacitor and the capacitor
memory is maximized to make the circuit less sensitive to value the sensitivity to charge injection and noise, there
the constant charge injection, and second, the signal de- is a tradeoff between speed and accuracy. The sampling
pendency of the charge injection is significantly reduced. switch on-resistance and the other nodes in the loop pro-
Furthermore, when the current memory is linear, even an duce non-dominant poles that affect the phase margin.
error linearly dependent on the signal can be tolerated,
since it only affects the signal amplitude. 3.2. Limitations
The current memory is shown in Figure 5. The gate- A potential problem in all circuits using time inter-
source voltage of the sampling switch (MS) is made vir- leaved parallelism is mismatch between the parallel cir-
tually constant by using bootstrapping. Neglecting the cuits. Such mismatch can produce an offset component
bulk effect, this makes the channel charge, as well as at half the clock frequency and a spectral image of the
the injection resulting from its release, constant. The ac- signal around it. Similar image can be generated by a
tual switch realization is based on the circuit presented clock skew in the half-rate clocks. Therefore any devia-
in [5]. Bootstrapping is also a reliable way to increase tion in the timing of the current switches should be min-
the gate voltage above the supply voltage without vio- imized. The current memory-based deglitcher, however,
lating the technology reliability specifications [6]. The does not rely on matching as long as charge injections of
maximum gate overdrive yields a low on-resistance with bootstrapped sampling switches and base currents of the
a small switch transistor, minimizing the nonlinear para- current switch pairs match. To reduce the skew error the
sitic capacitances, which are effectively in parallel with half-rate clocks for the current switch control circuit are
the memory capacitor. generated by dividing the full-rate clock with a carefully
Besides the channel charge, there is also charge re- matched synchronous divide-by-two circuit constructed
distribution in the gate overlap capacitance C Gol . Since, with a fully differential D-flipflop. For prototyping pur-
when entering the off-phase, the switch gate is pulled to poses a manual rise time control circuit was included in
a constant voltage, the resultant signal-dependent error the signal path for fine-tuning the duty cycle.
charge is CGol VG . When the capacitances are constant
(which is mostly true) and the current memory linear the 4. Experimental Results
error results in only a small change in signal amplitude.
If necessary, this error could be avoided by switching the The circuit was designed using a 0.35-m BiCMOS
gate to a voltage that is the V G properly buffered and (SiGe) technology and the chip (Figure 6) occupies a to-
level-shifted. tal 5.7 mm2 of silicon area. The majority of the area is
The high linearity of the current memory is based on consumed by the DAC current sources, the deglitcher be-
the fact that the transistor M1 (Figure 5) is biased in the ing only a small block on the lower right corner of the
481
IMEM POWER SPECTRUM @ 30 MS/s
10
Vref + M2 f
sig
0
A1 Q1
Clk − M3 −10
n1
−20
−60
Figure 5. BiCMOS current memory cell. 2fsig
−70
−80
−90
0 2 4 6 8 10 12 14 16
Frequency [MHz]
−10
−20
Figure 6. A microphotograph of the chip.
SIGNAL POWER [dBFS]
−30
−40
chip. Attention was paid to the power distribution issues fclk/2
and symmetry because any deviation in the ground volt- −50
482
DESIGN AND IMPLEMENTATION OF VIDEO DAC IN
0.13pm CMOS TECHNOLOGY
Cristiun Ionuqcu Dun Burdid
'Department of Applied Electronics and Intelligent Systems 2Department of Telecommunications
e-mail address: cionascuQetc.tuiasi.ro dburdia@etc.tuiasi.ro
Faculty of Electronics & Telecommunications
"Gh. Asachi" Technical University o f l a s i
ABSTRACT are high glitch energy even with very carefully design of switches
and large DNL (differential nonlinearity) error [Z].The second
This paper presents the design and implementation of video supposes a conversion of digital inputs into a special code named
digital-to-analog converter (DAC) in CMOS 0.13pm technology thermometer, where the number of the outputs is equal to the
on 8 bits resolution, 500 MS/s, with good performances, smaller decimal value of the input signal. The advantages of this
area low glitch energy, which makes it suitable for RAMDAC structure are good DNL and small dynamic switching errors. The
(Random Memory Access Digital to Analog Converter) mathematical description of current sources behavior for these
application. The proposed current steering DAC is a fully two StNCtUreS is:
thermometer architecture which assumes a smaller area for a
single DAC unit with a new deglitching circuit architecture u ( N )= K * L S B ( I ) (Binary-weighted)
added to improve the dynamic performances. In this way the I
interface between the digital part, which operates at low power
supply 1.2V. and the analog part operating at 3.3V supply is
made possible. AI) = 0
I
LSB (2) (Unary current sources)
a RAM circuit) in analog signal for VGA display. The most 8Digibl
important block is the DAC which has to operate at pixel clock Inputs
frequency with minimum distortions, to provide high current
scale and precise voltage full scale level compatible with
standards such RS-170, RS-343 [ 6 ] and standard load
terminations (75 Ohm). Its speed determines the rtsolution level
and corresponding refresh rate. Faster RAMDAC speeds. allow
more pixels per second to be placed on the screen. This result in
two important benefits: higher resolution allows the viewer to see
better and a faster refresh rate reduces flicker and eyestrain.
Most of conventional high-speed digital to analog converters
employ current steering architectures for their speed and
linearity. However, the current steering DAC's show static and 4b DAC MSB CELL
dynamic performance limitations due to process variation,
current mismatch, and glitch energy.
There are two well-known current steering architectures: binary-
weighted and unary decoded. In the first, every input signal
drives directly a switch that leads the current source hinary-
weighted to the output load. The drawbacks of this architecture Fig1 The block schematic ofRAMDAC (a single DAC.unit)
382
Another iechnique to increase the dynamic performances of DNL, INL (integral nonlinearity) errors. To get maximum
this DAC srchitecture is to make rising and falling edges of performances under these conditions, a schematic for the current
driving signals, especially for MSB switches, as equal as cell is shown in fig.5. All transistors used in this cell (except
possible, by taking into account the larger capacitance that h p to deglitching circuit) are 3.3V transistors, M2 and M3 represents
be driven by the outputs signals of the deglitching circuit for the switches, MO and M I the unary current source.
MSB switches in ccmparison with LSB switches (I_MSB=16 Many trade-offs have to be accomplished when designing
1-LSB). AS it is shown in fig4 the signals out and outn have a this block, for example the switches have to be large to lower the
lower crossing point with almost equal rising and falling edges, effect of their own parasitical resistance, but on the other hand,
more delay being added between them as compared to the signals this will increase glitches amplitude, due to the .
out-latch and outn-latch. clockfeedthrough (parasitical coupling through Cgs of MO and
MI for both edges of driving signals). The solution is to
minimize the sizes of switches and to increase the output
I
.i Jn
impedance for the current sources. In this work, for the current
.:ul sources a single cascode stage was preferred. There f e many
r
converters with a single transistor as the current source, using the
switch in saturation region to increase the overall output
impedance, especially for small full-scale currents.
In this design, a minimum possible size of the switches with
maximum possible impedance for current sources in case of a
large LSB current value (max 65.5 PA) was obtained.
Another important problem is matching of the unary current
sources. Using 131, the investigations show that the area of the
objects matching can characterize the variance of the stochastic
relative matching errors. As it was shown in most of the articles
related to matching behavior of CMOS transistors, the larger
transistors have a better performances in matching parameters.
So, to achieve a good matching, a special layout techniques was
developed (double common centroid): This was also
accomplished using the connection of several current sources in
parallel. Sizing of current sources is difficult because of limited
headroom (by the necessity of large full-scale currents). This
Fig.4 Output waveforms plotted according to fig.? limits the maximum size transistors and the limited output
impedance of current sources that has to be achieved gives
minimum limit. In fig.6 the transient simulations results for
IOOMhr update frequency with RC extracts is presented. The
glitch energy is about 6.5pV'sec.
~ 0 m Q s m n m ~ c o ~ w I o I p .
n"rInADml0lol
. __"- m
-s -- I 1.- (I T I
-.
3 83
3. LAYOUT CONSIDERATIONS I
384
Mismatch Modeling and Design of CMOS Current-
Steering Digital-to-Analog Converters
Ola Andersson
Reg no: LiTH-ISY-EX-3026
November 3, 1999
Mismatch Modeling and Design of CMOS Current-
Steering Digital-to-Analog Converters
Thesis for Degree of Master of Science
Department of Electrical Engineering
Division of Electronics Systems
Ola Andersson
Reg no: LiTH-ISY-EX-3026
November 3, 1999
DAC
LiTH-ISY-EX- 3026
URL för elektronisk version
Sammanfattning
Abstract
The requirements on today’s telecommunications equipment are very hard. Most of the signal processing
is done in the digital domain, but the information has to be transferred with analog signals, and therefore
analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) are crucial building blocks.
In this report we concentrate on the matching properties of current-steering DACs. Because transistor-
level simulations are time demanding, we need to approach the problems on a higher level of abstraction.
A mathematical model of so called linearly graded mismatch has been developed and analyzed for a par-
ticular DAC structure.
A DAC structure that will reduce the influence of linearly graded mismatch has been developed, and two
sets of DAC chips with different types of current switches has been implemented. Some design issues for
mixed mode applications in general, and especially DACs, are also discussed.
Nyckelord
95-11-01/lli
Keywords
DAC, matching, mixed mode design, current source, CMOS, analog design
Abstract
The requirements on today’s telecommunications equipment are very hard. Most of the
signal processing is done in the digital domain, but the information has to be transferred
with analog signals, and therefore analog-to-digital converters (ADCs) and digital-to-ana-
log converters (DACs) are crucial building blocks.
In this report we concentrate on the matching properties of current-steering DACs. Be-
cause transistor-level simulations are time demanding, we need to approach the problems
on a higher level of abstraction. A mathematical model of so called linearly graded mis-
match has been developed and analyzed for a particular DAC structure.
A DAC structure that will reduce the influence of linearly graded mismatch has been de-
veloped, and two sets of DAC chips with different types of current switches has been im-
plemented. Some design issues for mixed mode applications in general, and especially
DACs, are also discussed.
Acknowledgments
I would like to thank my supervisor Tekn.Lic. J. Jacob Wikner for his invaluable help in
introducing me to the world of data conversion, teaching me how to use all the IT-tools and
giving me useful remarks on my work.
My deepest gratitude to Ericsson Components AB and Ph.D. Gunnar Björklund for giving
me the opportunity to do this work.
Special thanks to my office colleagues Niklas Andersson and Pierre Dahlheim Lander,
without you two the spirit of the office would not have been on the same high level. I also
thank Elmo for never complaining.
Finally I thank my fiancé, Helena, for always supporting me.
Table of Contents
1 Introduction................................................................................... 1
1.1 Background........................................................................... 1
1.2 Brief Overview of the Work ................................................... 2
1.2.1 Background Research ................................................................. 2
1.2.2 High-Level Approach ................................................................... 2
1.2.3 Implementation ............................................................................ 2
i
3.5.1 Mismatch Modeling...................................................................... 15
3.5.2 DC Properties Arising from Linearly Graded Mismatch ............... 17
3.5.3 SFDR due to Linearly Graded Mismatch ..................................... 18
3.5.4 SNDR due to Linearly Graded Mismatch..................................... 21
3.5.5 Simulated Signal Spectras........................................................... 24
3.6 Conclusions .......................................................................... 24
Appendices
A Using MatLab for Behavioral-Level Simulations ........................ 39
A.1 Introduction ........................................................................... 39
A.2 Brief MatLab Overview ......................................................... 39
A.3 Creating DAC Models using MatLab .................................... 39
A.4 Creating Input and Output Signals for the DAC.................... 39
ii
A.5 Analyzing Signals in the Frequency Domain ........................ 40
A.6 M-file Listings........................................................................ 40
A.6.1 createDAC.m ............................................................................... 40
A.6.2 corr_choice.m .............................................................................. 41
A.6.3 INLest.m ...................................................................................... 41
A.6.4 DNLest.m..................................................................................... 42
iii
iv
Abbreviations
Notation Description
2 Variance of X
Var ( X ) = σ (X )
ADC Analog-to-digital converter
ADSL Assymmetric digital subscriber line
CFT Clock feedthrough
DAC Digital-to-analog converter
DNL Differential non-linearity
DSL Digital subscriber line
FFT Fast Fourier transform
HDSL High data rate digital subscriber line
INL Integral non-linearity
LSB Least significant bit
MSB Most significant bit
SFDR Spurious-free dynamic range
SNDR Signal-to-noise-and-distortion ratio
SNR Signal-to-noise ratio
VDSL Very high data rate digital subscriber line
v
Chapter
1
Introduction
1.1 BACKGROUND
With the number of Internet users growing at a very high rate, the need for better and fast-
er telecommunications equipment is increasing. Digital subscriber line (DSL) is one tech-
nology for bringing high bandwidth information over ordinary copper telephone lines [1].
The reason for wanting to use copper lines is that this infrastructure already exist. xDSL
refers to different variations of DSL, such as HDSL (high bit-rate DSL), ADSL (asymmet-
ric DSL) and VDSL (very high data rate DSL).
Most of the signal processing is performed digitally because this is more accurate than if
analog circuitry were to be used. However, digital data has to be transferred on a long dis-
tance line using only analog signals, and therefore DACs and analog-to-digital converters
(ADCs) are crucial building blocks, or even the bottle-necks in some applications.
There is good knowledge in how to make high-speed/low-resolution converters or low-
speed/high-resolution converters. However, what is needed today for telecom applications
is both high-speed/high-resolution converters.
1
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
1.2.3 Implementation
Two sets of DAC-chips were implemented. In chapter 4 some issues concerning the chip
design are discussed, and special care is taken to the knowledge achieved about mismatch
in the theoretical part of the work. Discussions about design methodologies and the Ca-
dence design tools used to implement the DACs have been placed in appendix B.
2
Chapter
2
Introduction to DACs
2.1 INTRODUCTION
The requirements on today’s telecommunication circuits are in the order of 25MHz band-
width and 14-bit resolution, or even more [1]. To meet these requirements, the design of
analog and mixed-signal circuits becomes very hard. One of the bottle-necks is the inter-
face between the digital signal processing circuits and the analog amplifiers and filters
sending and receiving the signal. The interface consists of the digital-to-analog converter
(DAC) and the analog-to-digital converter (ADC). In this report we concentrate on the
matching properties and design issues of current-steering DACs most often used in tele-
communications applications. In this chapter we will introduce the reader to the concept
of DACs.
3
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
The DNL value describes the deviation from an ideal step in the transition between the
codes X – 1 and X , and INL describes the overall deviation from the wanted output.
Output value 5
4 A real ( 3 )
A wanted ( 3 )
3
1
d0
0
0 1 2 3 4 5 6 7
Input code X
Deviations from the wanted transfer characteristics of the DAC can be divided into two
main groups, offset errors and gain errors. The offset error may be defined as the deviation
from the wanted minimum output, marked d 0 in Fig. 2-1. Two types of gain errors, linear
and non-linear, are illustrated in Fig. 2-2 below.
Because linear gain errors and offset errors do not affect the performance of the DAC, INL
and DNL are often calculated with respect to a best-fit straight line,
A best-fit ( X ) = kX + A offset . k and A offset are calculated using the least square method and
1800
1600
1400
linear gain error
Output value
1200
1000
800
the real outputs of the DAC, further discussed in [2]. Better values of INL and DNL than
in Eq.(2-1) and Eq.(2-2) are now given by:
4
Chapter 2 - "Introduction to DACs"
1.2
0.8
output
0.6 ts
0.4
0.2
0
0 5 10 15 20
time
tr
Fig. 2-3: Ideal step (dashed) and real step response (solid)
1. Rise time, t r
The time it takes for the output to go from 5% to 95% of it’s final value
2. Settling time, t s
The time it takes for the output to settle within 5% of it’s final value
The percentages given in the above definitions are often used, but they are not a standard,
so the reader should not be surprised if other values are given in the literature.
When we have relatively small code transitions, the output can be well approximated with
a linear system, and we have so called linear settling. However, for larger code transitions
the output driving capability is a limiting factor. We define the slew rate as
5
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
3. Slew rate, SR
SR = max dY (2-5)
dt
where Y is the output value.
The dynamic properties are strongly dependent on which codes are involved in the transi-
tion, and are therefore difficult to model. In this work we limit ourselves to model the in-
fluence of the static errors.
6
Chapter 2 - "Introduction to DACs"
−80
−100
−120
−140
where X rec(t) is the value of the reconstructed signal at the time t , X ( n ) is the value of the
sampled signal at sampling instant n , and we sum over all values of n .
The sinc function is defined as
sin ( πx )
sinc(x) = ------------------- (2-10)
πx
Since the reconstruction of Eq.(2-9) is not possible to implement, the sinc-function has to
be replaced by a function that does not have an infinite extension in time. Most often the
sinc is replaced by a square pulse, as shown in Fig. 2-5 where the sinc also is plotted. When
we refer to the ideal DAC in this report, we mean a DAC where the sinc has been replaced
by a square pulse, i.e., sample-and-hold elements are used at the output. This causes the
signal spectrum to be sinc weighted.
This definition of the ideal DAC does not really correspond to right meaning of the word
ideal, since the reconstruction of the signal is indeed not ideal. It is instead a description
of what we want to achieve when designing DACs.
7
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
time
Fig. 2-5: The sinc function (dashed) plotted together with the square pulse (solid) used instead.
∑ bk 2
k
A ( X ) = A ref ⋅ (2-11)
k=0
A ref in Eq.(2-11) is a reference value, often an electrical voltage, current or charge.
8
Chapter 2 - "Introduction to DACs"
a)
b)
I0=Iunit I1=2Iunit IN-1=2N-1Iunit
∆k
Ik
b0 b1 bN-1 bk
Iout Iout,k
Fig. 2-6: a) Current-steering DAC and b) unit current source with matching error ∆k.
where it is terminated over a resistive load (typically 50 Ohms). The N -bit DAC in Fig. 2-
6a) is constructed by using N binary-weighted current sources, I k . Each current source I k
utilizes 2 k parallel unit current sources, I unit . The bits b i choose which current sources to
connect to the output. b 0 is the least significant bit (LSB) and b N – 1 is the most significant
bit (MSB), hence the total output current for an ideal DAC is given by (compare with Eq.(2-
11))
9
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
10
Chapter
3
3.1 INTRODUCTION
Analog designs are often dependent on certain relations between devices used in the de-
sign. Examples of such relations can be that two transistors should be identical or that
there should be a certain relationship between their sizes. When fabricating the design,
those relations will not be fulfilled due to imperfections in the fabrication process, causing
so called matching errors. In this chapter we will describe the matching errors and how
they affect the performance of the current-steering DAC.
a) drain (D) b) S
gate (G)
bulk (B) G B
(substrate)
source (S) D
11
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
For an NMOS device the substrate is of p-type, whereas the source and the drain are heavi-
ly n-doped. For a PMOS device the substrate is of n-type, whereas the source and the drain
are heavily p-doped. A deeper discussion on the physics behind the metal oxide semicon-
ductor field effect transistor (MOSFET) can be found in e.g. [8], in this report we only
present the approximate relations between currents and voltages in a MOSFET.
1. Cut-off region ( v GS ≤ v T , n )
iD ≈ 0 (3-1)
2. Linear region ( v GS > v T , n ) and ( v DS ≤ v GS – v T , n )
µ 0, n C ox W
i D ≈ ------------------- ----- ( 2 ( v GS – v T , n ) – v DS )v DS ( 1 + λv DS ) (3-2)
2 L
3. Saturation region ( v GS > v T , n ) and ( v DS > v GS – v T , n )
µ 0, n C ox W 2
i D ≈ ------------------- ----- ( v GS – v T , n ) ( 1 + λv DS ) (3-3)
2 L
The threshold voltage, v T , n , can be determined using
v T , n = V T 0, n + γ ( 2 Φ F, n – v SB – 2 Φ F, n ) (3-4)
The mobility µ 0, n , the oxide capacitance C ox , V T 0, n , γ and Φ F, n are constants for the
specific process. The subscript n indicates that the constant is specific for an NMOS de-
vice. The channel length modulation factor λ is dependent on the transistor size, but is
often set to zero when less accuracy is needed. The width, W , and the channel length, L ,
of the transistor are indicated in Fig. 3-2. The drain current, i D , and the different voltages
are defined in Fig. 3-3a).
12
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
v SG
iD v DS v SB
v BS v SD
v GS iD
a) b)
Fig. 3-3: Definitions of voltages and drain current in a) an NMOS transistor and b) a PMOS
transistor.
Even though the linear relation between i D and v GS that is present in the linear region is
not present in the saturated region, transistors in analog designs are most often operated
in the saturated region. This is because the current is (almost) only dependent on one volt-
age, v GS , and therefore more easily controlled than then the current for a transistor in the
linear region. An exception to this is when the transistor is used as an analog switch, and
then operated in the linear or cut-off region.
1. cut-off region ( v SG ≤ v T , p )
iD ≈ 0 (3-5)
2. linear region ( v SG > v T , p ) and ( v SD ≤ v SG – v T , p )
µ 0, p C ox W
i D ≈ -------------------- ----- ( 2 ( v SG – v T , p ) – v SD )v SD ( 1 + λv SD ) (3-6)
2 L
3. saturation region ( v SG > v T , n ) and ( v SD > v SG – v T , p )
µ 0, p C ox W 2
i D ≈ -------------------- ----- ( v SG – v T , p ) ( 1 + λv SD ) (3-7)
2 L
The threshold voltage for the PMOS transistor, v T , p , is determined by
v T , p = V T 0, p + γ ( 2 Φ F, p – v SB – 2 Φ F, p ) (3-8)
The different process parameters are similar to those for the NMOS transistor. Here the
subscript p indicates that the constant is specific for the PMOS device. The drain current,
i D , and the different voltages are defined in Fig. 3-3b)
3.2.3 Small-Signal Equivalents
When analyzing the small-signal response of CMOS circuits, linearized models, small-sig-
nal equivalents, of the MOSFET are used. A simple small-signal equivalent of an NMOS
is shown in Fig. 3-4 below.
The small-signal parameters in Fig. 3-4 are defined as
13
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
G + D
v gs 1
g m v gs r ds = -------
g ds
-
S
∂i D
gm = (3-9)
∂ v GS
Q
∂i D
g ds = (3-10)
∂ v DS
Q
The subscript Q denotes that the partial derivatives are evaluated in the quiescent (DC
operation) point.
In the saturated region, Eq.(3-9) and Eq.(3-10) become
W W
g m ≈ µ 0, n C ox ----- ( v GS – v T , n ) = 2µ 0, n C ox ----- i D (3-11)
L L
and
µ 0, n C ox W 2
g ds = λ ------------------- ----- ( v GS – v T , n ) ≈ λi D (3-12)
2 L
A further analysis, given in [9], gives the following result for the variance of the parameter
mismatch:
2
2 A 2 2
σ ( ∆P ) = -------P- + S p D x (3-13)
WL
In the above equation W and L are the transistor width and length, D x is the distance
between the compared transistors, and A P and S P are constants (for the particular wafer)
describing the short- and long-distance variations of parameter P respectively. From
Eq.(3-13) we can see that for good matching properties we should place the transistors as
close together as possible, and keep the product WL as large as possible.
14
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
a) b)
Fig. 3-5: a) Layout reducing gradient matching errors and b) straight forward layout
uniformly over the array of current sources. Even though every unit source has errors as
described previously, this type of layout tends to cancelate those errors slightly. Fig. 3-5b)
shows a more straight forward type of layout. Since the unit sources for each bit are placed
very close to each other their errors are very much correlated and the total error for a par-
ticular bit can be quite large. However, this kind of layout has some advantages concern-
ing, e.g., wire routing.
15
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
a) b)
i(1,1)
i(-1,1)
I5
I4
i(-2M-1,-2N-M-1)
i(1,-1) I3
y dummy
I2 I1 I0
x source
Fig. 3-6: a) Array of unit current sources with defined directions and positions.
b) Assignment of unit sources to specific bits for a 6-bit DAC.
We model the unit current sources in the same way as the binary weighted sources, hence
a nominal current source in parallel with an error source (compare with Eq.(2-17))
i ( a, b ) = I unit + δ ( a, b ) (3-14)
Using the notation above we have:
I = 2k ⋅ I
k unit
(3-15)
∆k =
∑ δ ( a, b )
( a, b ):i ( a, b ) ∈ I k
Investigations of CMOS current sources [12] show that the deterministic part of the mis-
match often show a linear distribution over the chip, i.e. the geometrical dependence of the
current sources can be well approximated with it’s first order Taylor expansion.
Neglecting the short distance variations and assuming that the matching errors are lin-
early distributed over the chip area (current source array) we can express the error sources
as
M – 1 – 2 k + 1 + 2 k – 1 ) – ( k ⋅ ( 2 N – M – 1 – 2 – 1 ) ) ) ⋅ 2 k, 0 ≤ k ≤ M – 1
(kx ⋅ (2 y
∆k =
– k y ⋅ ( 2 N – M – 1 – 2 k – M + 1 + 2 k – M – 1 ) ⋅ 2 k, M≤k≤N–1
(3-17)
From Eq.(3-17) we can see that the gradient in the x-direction, as could be expected, only
k
affects the M LSBs. Due to the 2 factor in Eq.(3-17) the largest errors will occur for the
MSBs, provided that k x is not very large compared to k y , which probably is a valid as-
sumption.
16
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
a) DNL for an 8−bit DAC with M=4, ky=0.001 and kx=0 b) INL for an 8−bit DAC with M=4, ky=0.001 and kx=0
1 0.5
0.9 0.4
0.8 0.3
0.7 0.2
0.6 0.1
DNL
INL
0.5 0
0.4 −0.1
0.3 −0.2
0.2 −0.3
0.1 −0.4
0 −0.5
50 100 150 200 250 0 50 100 150 200 250
Input word Input word
Fig. 3-7: Typical a) DNL and b) INL plot for an 8-bit DAC with linearly
graded matching errors only in the y-direction.
a) DNL for an 8−bit DAC with M=4, ky=0.001 and kx=0.001 b) INL for an 8−bit DAC with M=4, ky=0.001 and kx=0.001
1 0.5
0.9 0.4
0.8 0.3
0.7 0.2
0.6 0.1
DNL
0.5
INL
0
0.4
−0.1
0.3
−0.2
0.2
−0.3
0.1
−0.4
0
−0.5
50 100 150 200 250 0 50 100 150 200 250
Input word Input word
Fig. 3-8: Typical a) DNL and b) INL plot for an 8-bit DAC with linearly
graded matching errors in both the x and y direction.
In Fig. 3-8 we show the simulated DNL and INL when introducing a gradient error in the
x-direction as well. The gradients are given by k x = k y = 0.001 [ LSB ⁄ unit source ] . As ex-
pected these plots are very similar to those in Fig. 3-7. Even though we can see that the M
LSBs no longer are mutually matched, the most prominent errors are still those from
17
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
f0
I out(n) ≈ 2 N – 1 ⋅ sin 2π ⋅ ----- ⋅ n + e ( n ) + I DC (3-18)
fs
where f 0 is the signal frequency, f s is the sample or update frequency, I DC is the signal’s
DC level, n is the sequence index, and e ( n ) is the error signal due to the linearly graded
mismatch. From Eq.(2-16) we have that the error signal can be written as
N–1
e(n) = ∑ bk(n) ⋅ ∆k (3-19)
k=0
It can be shown that
N–2
∑ ∆k ≈ –∆ N – 1 (3-20)
k=0
Using Eq.(3-20) we can calculate the mismatch error for four different inputs
0, X = [ 0, 0, …, 0, 0 ] or [ 1, 1, …, 1, 1 ]
–∆
e = N – 1, X = [ 0, 1, …, 1, 1 ] (3-21)
∆ N – 1, X = [ 1, 0, …, 0, 0 ]
Using this information we approximate the error signal with a piece-wise linear signal,
ẽ(n) , as shown in Fig. 3-9. Making a Fourier series expansion on the approximated error
signal gives us
2
−0.5 ẽ(n)
−1 –∆ N – 1
−1.5
−2
0 1 2 3 4 5 6
Fig. 3-9: Input signal (dashed) and approximative error signal (solid) plotted in different scales.
18
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
∞ 0, k even
f
ẽ ( n ) = ∑ A k sin 2πk -----n where A k =
0
(3-22)
∆ N – 1 ------ 1 – ------ , k odd
fs 4 2
k=1 πk πk
Besides A 1 , the largest coefficient A k is given by the 3rd harmonic, A 3 . Also, since we
have applied a full-scale sinusoid at the input, the signal amplitude is approximately
N–1
A signal ≈ 2 (3-23)
Combining Eq.(3-22) and Eq.(3-23) gives us the spurious-free dynamic range (SFDR) as
2
A signal ⁄ 2 2 2 2( N – 1)
SFDR ≈ ---------------------- 9π 2
- = ----------------------- ⋅ ------------------
- (3-24)
2 4 ( 3π – 2 ) 2
A3 ⁄ 2 ∆N – 1
Using Eq.(3-17) for ∆ N – 1 in Eq.(3-24) we get
2 2 N – 1 M + 1 2 2 2 2M
9π 2 2
SFDR = ----------------------- ⋅ ---------------------------
–2 9π 2
⋅ k y = --------------- ⋅ --------
–2
- ⋅ ky
4 ( 3π – 2 ) 2 ( N – 1 )
(3-25)
3π – 2 2N
2 2
Rewriting Eq.(3-25) in dB yields
90
80
M=12
SFDR [dB]
70
60
50
40
M=6
30
12 13 14 15 16 17
Number of bits
Fig. 3-10: SFDR as a function of N for different values of M when ky=0.0001 and kx=0.
19
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
80
SFDR [dB]
70
N=17
60
50
40
30
6 7 8 9 10 11 12
Number of bits in one row of the array
Fig. 3-11: SFDR as a function of M for different values of N when ky=0.0001 and kx=0.
From those two simulations we can get the following “empiric” expression for SFDR:
SFDR ≈ K + 6M – 6N dB (3-27)
This means that we have a verification of Eq.(3-26) as far as to the dependence of N and
M . K is a function of k y not yet determined.
In Fig. 3-12 the calculated value of SFDR is plotted as a function of k y together with sim-
SFDR as a function ky, kx=0
150
100
SFDR [dB]
65.5
50 65
64.5
64
−4
10
0 −8 −7 −6 −5 −4 −3 −2
10 10 10 10 10 10 10
Gradient in the y−direction [LSB/(unit source)]
Fig. 3-12: Calculated (solid) and simulated (dashed) values of SFDR as a function of ky for a
14-bit DAC with M=8.
ulated values for a 14-bit DAC with an array width of M = 8 . We see a close agreement
between calculated and simulated values for larger gradients, for smaller gradients the
simulated values are limited by the quantization noise, which is dependent on the FFT
length used in the simulations.
20
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
N–1
Var ( e(n) ) = ∑ Var ( e k(n) ) + ∑ Cov ( e i(n), e k(n) ) (3-28)
k = 0 ( i, k ):i ≠ k
When applying a full-scale sinusoid, the dominating covariance term, appearing twice in
the sum above, is
2
∆N – 1
Cov ( e N – 1(n), e N – 2(n) ) = – -------------
- (3-29)
48
The error signals, e k ( n ) , have the following probability functions:
1
--- x = 0 or ∆ k
P ( ek ( n ) = x ) = 2 (3-30)
0 otherwise
Since we have the probability functions we can calculate the variance for each error signal,
yielding
2 2 2
1 ∆ ∆ ∆
Var ( e k ( n ) ) = --- ⋅ -----2k- + -----2k- = -----k- (3-31)
2 2 2 4
( 2 N – M – 1 – 2 k – M + 1 + 2 k – M – 1 )2 M + 1 k
-2 M ≤ k ≤ N – 1
----------------------------------------------------------------------------------------------
2( N – 1)
∆k 2
-------------- = (3-32)
∆N – 1 k ( 2 N – M – 1 – 2 –1 )
– 2 ---------------------------------------
2( N – 1)
- 0≤k≤M–1
2
Some algebraic manipulations on the above expressions give us:
21
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
N–1 N–M
M – Na –1 1
∑
k–N N–M
a = a ------------------------ ≈ { if a >>1 } ≈ ------------ (3-34)
a–1 a–1
k=M
This gives us the expression
N–1
∆ k 2 16 96 144
-------------
∑
128
- ≈ ------ – ------ + --------- = --------- (3-35)
∆ N – 1 3 7 15 105
k=M
In a similar way for the LSBs we get
2
M–1
∆k 2
------------- 2 N – 2 M 4 M – 1 1 24
M
16 M – N
∑ ∆ N – 1- = 16 -------------------
2N
---------------- ≈ 16 -----N- ------- = ------ 4
4–1 3 3
(3-36)
k=0 2 2
Using these relations we arrive at
N–1 N–1
∆k 2
-------------
M–N
- ≈ ∆ N – 1 ------------------------------------
1 2 32 + 4 140
∑ ∑
2
Var ( e k(n) ) = --- ∆ N – 1 (3-37)
4 ∆ N – 1 105
k=0 k=0
Now we can rewrite Eq.(3-28) as
M–N
32 + 140 ⋅ 4
Var ( e(n) ) ≈ ∆ N – 1 ⋅ ----------------------------------------- – ------
2 1
(3-38)
105 24
We are now ready to express the power of the noise and distortion, P ND , as
2
A
P ND = Var ( e(n) ) + P q – -----1- (3-39)
2
In Eq.(3-39) we have included P q , the power of the quantization noise, and also subtracted
the power of the error signal that appears at the signal frequency f 0 (compare with Eq.(3-
22)). It is known that P q = 1 ⁄ 12 , and from Eq.(3-22) the value of A 1 follows. We can thus
write Eq.(3-39) as
32 + 140 ⋅ 4 M – N 1 2 2 8
P ND = ------ + ∆ N – 1 ⋅ ----------------------------------------- – ------ – 1 – --- ⋅ ----2-
1 2
(3-40)
12 105 24 π π
22
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
80
70
60
SNDR [dB]
50 61
60.95
40
60.9
30 60.85
60.8
20 10
−4
10
0 −7 −6 −5 −4 −3 −2
10 10 10 10 10 10
Gradient in the y−direction [LSB/(unit source)]
Fig. 3-13: Calculated (solid) and simulated (dashed) values of SNDR as a function of ky for a
14-bit DAC with M=8.
23
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
In Fig. 3-14 the SNDR approximation of Eq.(3-47) is plotted together with simulated val-
ues, as in Fig. 3-13. As we can see from the simulations, Eq.(3-41) gives us a much better
approximation of the SNDR than Eq.(3-47) did, on the other hand the latter was much eas-
ier achieved and gives acceptably good results.
80
70
60
SNDR [dB]
50
61
40
60.5
30
60
20
59.5
10 −4
10
0 −7 −6 −5 −4 −3 −2
10 10 10 10 10 10
Gradient in the y−direction [LSB/(unit source)]
Fig. 3-14: Approximated (solid) and simulated (dashed) values of SNDR as a function of ky for
a 14-bit DAC with M=8.
3.6 CONCLUSIONS
In this chapter we have discussed mismatch of CMOS current sources and presented a
study of a special case of matching errors, linearly graded mismatch. The study was car-
ried out for one particular way of choosing current sources from an array in a CMOS DAC,
a technique known not to be good in suppressing linearly graded matching errors. Howev-
er, the work presented can be seen as a proposed approach to the concept of linearly graded
matching errors that can be applied on other layout styles. It is desirable to keep the layout
as simple as possible in order to keep the amount of time needed to design the DAC as low
as possible. Given expected process parameter variations from the vendor one can, on a
high abstraction level, find how complex layout style is needed to reach the wanted perfor-
mance, e.g. the layout style of Fig. 3-5a) will ideally reduce the influence of linearly graded
mismatch completely, but it is much more complex to design than the layout style of Fig.
3-5b). If the latter is sufficient for the wanted application, it is preferred that it is used in-
stead.
24
Chapter 3 - "Modeling of the Influence of Linearly Graded Mismatch"
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Singel−tone output from 14−bit DAC, M=8, kx=0, ky=0.0001
0
Signal Power [dB]
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Singel−tone output from 14−bit DAC, M=8, kx=0, ky=0.001
0
Signal Power [dB]
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalized frequency
Fig. 3-15: Signal spectra for different values of ky, when kx=0.
Singel−tone output from 14−bit DAC, M=8, kx=0.001, ky=0
0
Signal Power [dB]
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Singel−tone output from 14−bit DAC, M=8, kx=0.001, ky=0.0001
0
Signal Power [dB]
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Singel−tone output from 14−bit DAC, M=8, kx=0.001, ky=0.001
0
Signal Power [dB]
−50
−100
−150
−200
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalized frequency
Fig. 3-16: Signal spectra for different values of ky, when kx=0.001.
25
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
To get a more realistic model of the DAC one needs to introduce more non-idealities than
just the linearly graded mismatch, e.g. short distance variations, wire resistance, finite
output impedance of current sources and circuit noise. Modeling of the mentioned non-ide-
alities are discussed in [4]. A suggestion for future work is to combine those properties into
a DAC model to help DAC designers optimize their design on a high level of abstraction.
26
Chapter
4
Implementation of Current-Steering
DACs
4.1 INTRODUCTION
In the previous chapters the current-steering DAC has been described in a relatively ab-
stract way concerning mismatch and geometrical structure, but we have not described how
the DAC is really implemented. In this chapter we describe in more detail the implemen-
tation of two 14-bit DACs with two different types of current switches, and where the cur-
rent sources are placed to decrease the influence of linearly graded mismatch as described
in chapter 3.
The DACs were implemented in a 0.35µm standard digital CMOS process with three met-
al layers and two poly layers. The DACs were designed to operate at sampling frequencies
up to 88MHz , using a supply voltage of 3.3 V .
4.2.1 Segmentation
So far we have considered DAC structures where each bit is represented by a binary-
weighted current source which is turned on and off when the bit is 1 and 0 respectively. A
disadvantage with this structure is a phenomenon called glitches, unwanted current peaks
on the output. Glitches occur when some current sources are switched on while others are
switched off. The worst case is the transition between 011…11 and 100…00 . During this
transition we might for a moment have the incorrect input 111…11 (or some other incor-
rect input) due to non-perfect timing of the switching signals.
To avoid glitches we can use segmentation, a transformation of the binary coded input code
to thermometer code. This transformation is illustrated for a 3-bit binary input in Table 4-
1 below. We can see that during a code transition using thermometer code it will never oc-
cur that some sources are switched on while others are switched off, which means that we
will have less glitches.
With an N -bit binary word we will have a corresponding thermometer coded word of
2 N – 1 bits. When we have N = 14 , it is not suitable to have full segmentation, because
the thermometer coded word would then have 2 14 – 1 = 16383 bits. However, to avoid
large glitches we can use segmentation on some of the MSBs. In this implementation we
27
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
binary thermometer
b2b1b0 s7s6s5s4s3s2s1
000 0000000
001 0000001
010 0000011
011 0000111
100 0001111
101 0011111
110 0111111
111 1111111
Table 4-1: Example of segmentation
s31
s30
I8
I7 s2
I6 s1
I5
I4 non-segmented LSBs
I3 s1
I2 s2
I1
I0
s30
s31
a) b)
Fig. 4-1: Assignment of unit sources a) for the 9 LSBs and b) for the 5 segmented MSBs.
28
Chapter 4 - "Implementation of Current-Steering DACs"
in each row, where also the switches are placed, in order to keep the wire resistance as low
as possible. The unused unit sources (white space in Fig. 4-1) are used for biasing the other
unit sources, as described in section 4.3.2. In order to decrease the influence of linearly
graded mismatch it is preferred that the sources are chosen in the middle of each row, but
then the wire resistance will increase, and a trade off between those two properties has to
be made.
The sources for the thermometer coded bits are chosen in a way developed for reducing the
influence of linearly graded mismatch, as shown in Fig. 4-1b). Each of those sources use
two rows of unit sources, one from the upper half and one from the lower half of the array
equally spaced from the center in order to cancellate the gradient in the y-direction. The
gradient in the x-direction is already taken care of by the fact that we use an integer num-
ber of rows. With exactly linearly graded mismatch this arrangement would totally erase
the mismatch for those bits. However, the mismatch is probably not exactly linearly grad-
ed, and therefore we choose the thermometer bits most often turned off (s31, s30, ...) to con-
trol the sources furthest away from the center of the array, since those are the sources that
most likely have the largest mismatch.
1
I unit = ---------------------------
14
≈ 1.22 µA (4-1)
50 ( 2 – 1 )
In order to achieve high resolution of the DAC, the output resistance of the current sources
must be very large.. The total output resistance is a function of the input X which can be
written as
r out
R out = --------
- (4-2)
X
where r out is the output resistance of one unit source. Eq.(4-2) is motivated by the fact that
we have connected X unit sources in parallel.
In Fig. 4-2 the implementation of two types of current sources are shown. The single tran-
sistor current source in Fig. 4-2a) is the simplest possible current source, but the output
resistance of this current source, r out = 1 ⁄ g ds, 1 , is not large enough to reach high resolu-
tion. In Fig. 4-2b) we have included a cascode transistor M2 which increases the output
resistance to
g m, 2 1
r out = ------------ ------------ (4-3)
g ds, 2 g ds, 1
3
The increase of the output resistance is about a factor 10 . In order to further increase the
output resistance one could introduce even more cascode transistors, but then we will have
a larger voltage drop over the current source, which would decrease the possible output
swing of the DAC. For this implementation we choose the single cascode current source of
Fig. 4-2b), and reached an output resistance of r out ≈ 150 GΩ for a unit source.
29
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
vbias M1
vbias M1
M2
vcascode
iout
iout
a) b)
Fig. 4-2: a) Single transistor current source and b) single cascode current source.
4.3.2 Biasing
To bias the single cascode current source we use some of the unit sources in the array not
connected to any of the switches to form a wide-swing current mirror [2], as shown in Fig.
4-3. 1024 unit sources are used for the bias circuitry, and the current i ref is therefore
i ref = 1024 ⋅ 1.22µA ≈ 1.25mA . The cascode voltage, v cascode , is chosen to give as large
output swing as possible, keeping all transistors in the saturated region. Both i ref and
v cascode are produced off-chip.
vcascode
unit source
iref
Fig. 4-3: Biasing the unit sources with a wide-swing current mirror.
30
Chapter 4 - "Implementation of Current-Steering DACs"
Q Q
I+ I-
put giving rise to an unwanted current peak. To reduce CFT we can use the so called dum-
my switch of Fig. 4-5b), the dummy transistor works as a capacitor storing the channel
charge while the switching transistor is turned off, and returning it when the switching
transistor is turned back on again. In order to store the right amount of charge, the dummy
transistor has the same length, L , as the switching transistor, but only half the width, W .
The switching signals Q and Q are further described in section 4.3.5. In this implemen-
tation we have designed two chips, one using the simple switch and another using the
dummy switch, in order to find which one works best.
switching
switching transistors
transistors
Q Q
Q Q
Q Q
I+ I-
I+ dummy I-
transistors
a)
b)
Fig. 4-5: a) simple differential CMOS switch b) differential CMOS dummy switch.
31
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
≥1 s1
2-3-encoding
s1
b0 b0
s2 ≥1 s2
b0 b1
≥1 s1 b1 s3
b1 s2 ≥1 s3
& s3 b2 s4
b)
a) s5
&
& s6
& s7
≥1 s1
N-(2N-1)-encoding
b0 b0 s1
bN-1 bN-1 s 2N – 1
≥1 s 2N – 1
bN s 2N
c)
& s 2N + 1
& s 2N + 1 – 1
When the number of bits to encode is large, the depth of this encoding circuitry is also
large. For every additional bit to encode, the depth is increased with one gate, or in an
actual implementation with two gates since an or and an and gate is implemented as a nor
and a nand gate with a following inverter. This means that for a large number of bits we
will have to reduce the clock frequency. To solve this problem we use pipelining, every
stage of the circuitry is followed by a D-flip-flop, and therefore the signal only has to prop-
agate through two gates on every clock cycle. Pipelining introduces a delay in the signal,
so we have to delay the binary coded LSBs to the same extent in order not to have a mal-
function of the DAC.
32
Chapter 4 - "Implementation of Current-Steering DACs"
Q Q
Those switching signals could be achieved using the circuitry in Fig. 4-8. The way the nand
gates are connected ensures that the outputs are never low at the same time. In order for
one of them to go low, the other has to be high.
D-flip-flop
bit input set
D ≥1 & Q
clk out
& Q
With the set signal high, Q is high and Q is low, and with the set signal low, Q follows
the output of the D-flip-flop.
33
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
analog parts
S
S W
31 bits
E I
thermo-
G T
meter
M C
coded
5 MSBs E H
N
T S
A I
G S UNIT SOURCE
T W
I N
I ARRAY
O A
L T
N C
G H
9 LSBs D E E
9 LSBs
E N S
L E
A R
Y A
T
O
digital parts R
S
34
Chapter 4 - "Implementation of Current-Steering DACs"
vdd connection
substrate connections to
improve performance
gate of M1
gate of M2
output terminal
for the clock to drive. It is also important for the function of the DAC that the clock signals
to the different flip-flops has approximately the same delay. Because of these two reasons
the clock is distributed over the chip area with a tree structure as shown in Fig. 4-11 below.
clock signal in
Fig. 4-11: Example of clock distribution for better buffering and equal clock delays
Each inverter in the last stage will only have a small load to drive. If the wires between
two stages all have the same length, the delay will also be the same for all clock signals
from the output of the last stage.
Common to all types of pads are the (relatively) large metal plates where the bonding wires
are attached. The first three types basically consist of a wire connected to the appropriate
35
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
place, whereas the digital input pad is terminated with a relatively large buffer, as sche-
matically shown in Fig. 4-14, in order to have well defined signals on the chip and good
driving capability of the pad. The input pad also includes protective diodes and resistance
to ensure that the input voltage of the buffer does not reach too high or too low values, also
shown in Fig. 4-14.
To cope with the large current peaks that might occur, capacitors with reservoirs of charg-
es are connected between VDD and GND where there is chip area available. The digital
part of the circuit is by no means optimized in size. It could be made much smaller, but
since there is a lot of chip area available, some distance is kept between building blocks to
simplify routing. The total size of the chip is approximately 2.7 × 2.7 mm, containing al-
most 45 000 transistors. Some chip data is given in Table 4-2.
analog part
digital part
capacitors
2.7 mm
Fig. 4-12: Plot of the implemented DAC, including the pad frame
36
Chapter 4 - "Implementation of Current-Steering DACs"
Buffer of inverters
PAD
protection
circuitry
Property Description
Supply voltage 3.3V
Area ( 1.7mm × 1.2mm ) core, ( 2.7mm × 2.7mm ) with pads
Transistor count ~ 45000
The specified signal bandwidth for the DAC is 11 MHz , which implies a sample frequency
of at least 22 MHz . To improve the SNR we can use oversampling [2], i.e., we use a larger
sample frequency than 22 MHz to distribute the quantization noise over a larger frequen-
cy range. The noise power appearing at frequencies above the signal band is then removed
with an analog low-pass filter. When using oversampling we operate the DAC at 44 MHz
or 88 MHz .
37
"Mismatch Modeling and Design of CMOS Current-Steering Digital-to-Analog Converters"
38
Appendix A
Using MatLab for Behavioral-Level
Simulations
A.1 INTRODUCTION
Due to the long simulation time, it is almost impossible to carry out transistor-level simu-
lations of DACs in order to characterize the influence of mismatch. Therefore it is neces-
sary to create behavioral-level models for the simulations. In this appendix we present how
such simulations can be carried out using MatLab.
39
Appendix A - "Using MatLab for Behavioral-Level Simulations"
1 1 1
b N – 1 b N – 2 … b0
2 2 2
1 2 b N – 1 b N – 2 … b0
[a , a , …] ↔ (A-1)
. . . .
. . . .
. . . .
i i
a denotes the (wanted) analog value at sample instant i , and b k is the k :th bit of the cor-
responding digital number, either 0 or 1 . This transformation is done using the M-file
a2d2 [15], referred to in the code listings in section A.6.
The output signal of the DAC can now be achieved using simple matrix multiplication,
which is easily done with MatLab.
1 1 1
b N – 1 b N – 2 … b0
2 2 2
1 2 b N – 1 b N – 2 … b0 T
output signal = [ ã , ã , … ] = × [ I N – 1, …, I 0 ] (A-2)
. . . .
. . . .
. . . .
i i
Note that, even for an ideal DAC, ã and a will differ, at least for some values of i , since
the transformation in Eq.(A-1) means a loss of accuracy.
A.6.1 createDAC.m
This function is used to produce a matrix containing the values of the different unit sources
from an array with linearly graded matching errors.
function [current_sources] = createDAC(NOB, kx, ky, NOx, curr_nom)
% current_sources = createDAC(NOB, kx, ky, NOx, curr_nom)
%
% createDAC returns an array of unit current sources with the nominal value
% curr_nom and a determinstic graded error.
% NOB is the number of bits to be realized, NOx is the number of
% current sources in the x-direction (no. of columns in the array), kx and
% ky is the mismatch [LSB/(unit source)] in the x and y direction
% respectively.
40
Appendix A - "Using MatLab for Behavioral-Level Simulations"
% find the needed number of current sources in the y-direction (rows in the
% array)
NOy = ceil(2^NOB/NOx);
A.6.2 corr_choice.m
This function picks unit sources from an array given by createDAC.m and sums the values
into binary weighted current sources. The unit sources are chosen in a straight forward
manor described in section 3.5.
function [sources, currents] = corr_choice(NOB, kx, ky, NOx, Ilsb);
currents = zeros(1,NOB);
sources = createDAC(NOB, kx, -ky, NOx, Ilsb);
x=0;
y=1;
for i=NOB:-1:1
for j=1:2^(i-1)
x=x+1;
if x > NOx
x = 1;
y =y+1;
end
currents(NOB+1-i) = currents(NOB+1-i)+sources(y,x);
end
end
A.6.3 INLest.m
This function estimates the INL with respect to a best-fit straight line for each digital in-
put word and also returns a vector of input words for easy visualization with the MatLab
plot function.
41
Appendix A - "Using MatLab for Behavioral-Level Simulations"
A.6.4 DNLest.m
This function estimates the DNL with respect to a best-fit straight line for each digital in-
put word and also returns a vector of input words for easy visualization with the MatLab
plot function.
function [DNL, inputword]= DNLest(current_sources)
42
Appendix A - "Using MatLab for Behavioral-Level Simulations"
43
Appendix A - "Using MatLab for Behavioral-Level Simulations"
44
Appendix B
Integrated Circuit Design with
Cadence
B.1 INTRODUCTION
In this appendix we describe the different steps of integrated circuit (IC) design. The tools
used are products of Cadence Design Systems Inc., which are assumed to be the most com-
mon tools when designing analog or mixed mode applications [16]. The design method de-
scribed is so called full-custom design, i.e. all transistors and wires are placed manually,
but there are also, at least for digital designs, a few high level design methods, using for
example VHDL, which might be preferred when larger designs are to be implemented.
Since IC is a large field, this appendix only serves as a brief presentation of the subject.
For a more extensive presentation we refer to literature treating the subject, i.e. [17] or
[18].
Schematic
Simulations
Layout LVS
DRC
Extraction
custom VLSI design. As seen from Fig. B-1 there are three main design steps
1. Design of the circuit on a schematic level.
2. Circuit layout.
3. Netlist extraction from the circuit layout.
When designing on schematic level the design is iterated until simulations show that the
circuit works satisfactory.
45
Appendix B - "Integrated Circuit Design with Cadence"
When the schematic is completed, a layout of the circuit has to be made, i.e., transistors,
wires etc. have to be placed on the silicon surface. Design rule checks (DRCs) are carried
out on the layout to ensure that it is possible to manufacture the layout.
A netlist is extracted from the layout. This netlist is compared with the schematic netlist
during a layout-versus-schematic check (LVS). If the netlists match, the layout is finished,
otherwise the layout has to be corrected.
The tools used to carry out the steps described in this section will be briefly presented in
the following sections.
a) b)
The simulations also lets you sweep different parameters, e.g. transistor sizes and bias
voltages, in order to find the best possible values of those parameters.
The output of a transient analysis for the testbench of Fig. B-3 is shown in Fig. B-4.
46
Appendix B - "Integrated Circuit Design with Cadence"
Fig. B-4: Input (lower curve) and output (upper curve) for a transient analysis of a CMOS
inverter.
47
Appendix B - "Integrated Circuit Design with Cadence"
48
Appendix C
Contribution to the NORCHIP
Conference, Nov 1999
49
Appendix C - "Contribution to the NORCHIP Conference, Nov 1999"
50
Modeling of the Influence of Graded Element
Matching Errors in Current-Steering DACs
K. Ola Andersson and J. Jacob Wikner
{olaa, jacobw}@isy.liu.se, Phone: +46-13-281721, Fax: +46-13-139282
Microelectronics Research Center, Ericsson Components AB
Dept. of E.E., Linköpings universitet, SE-581 83 Linköping, Sweden
Abstract — In analog and mixed-mode circuits the matching between circuit elements is cru-
cial. For example, in binary encoded digital-to-analog converters (DACs) the matching
between different bit weights can set the limit on the performance. Related to earlier work
modeling the influence of stochastic matching, the influence of graded element matching errors
on the performance of current-steering DACs is shown. Presented are calculated results that
correlate very well with simulated results. As performance measures we use both static mea-
sures as DNL and INL as well as frequency domain parameters as SNDR and SFDR. This dis-
cussion can also be applied to other DAC structures, for example switched-capacitor.
1. INTRODUCTION
The requirements on today’s telecommunication circuits are in the order of 25MHz bandwidth
and 14-bit resolution or even more [1]. A crucial bottle-neck is the interface between the digital
signal processing circuits and the transmitting/receiving amplifiers and filters. In the interface
we also have the digital-to-analog converter (DAC) and analog-to-digital converter (ADC).
In this paper we focus on the DAC, and especially the binary weighted current-steering DAC as
presented in Sec. 2 and in [2, 3, 4]. In binary weighted structures, the matching error is more or
less setting the performance of the DAC. The matching errors are both of stochastic nature and
determined by gradients in for example the oxide thickness and voltage drops over supply
wires [5, 6]. In Sec. 3 we discuss the modeling of the mismatch error for a certain DAC topol-
ogy. This type of modeling has previously been reported in for example [4]. The DAC analyzed
in this paper has a very simple structure known to be bad in suppressing deterministic mis-
match, but the approach to deterministic mismatch that we present can be applied to other
types of DACs and help us comparing different layout styles.
Mostly, wideband DACs are characterized by their behavior in the frequency domain [1]. Espe-
cially we use the signal-to-noise-and-distortion ratio (SNDR) and the spurious-free dynamic
range (SFDR). Some important formulas determining the SFDR and SNDR have been derived
and they are presented in Sec. 5. In Sec. 4 we also briefly address the static measures in means
of the differential nonlinearity (DNL) and the integral nonlinearity (INL).
51
a) b)
I0=Iunit I1=2Iunit IN-1=2N-1Iunit
∆k
Ik=2k-1. Iunit
b0 b1 bN-1 bk
Iout Iout,k
Figure 1: a) Current-steering DAC and b) unit current source with matching error ∆k.
M N–M
The unit sources are arranged in an array with 2 sources in the x-direction and 2
sources in the y-direction, indexed as in Fig. 2a). Although it is not a very good design strategy,
we now can form the binary weighted current sources as shown in the example of a 6-bit DAC
in Fig. 2b), where M = 3 .
i(1,1)
i(-1,1)
I5
a) b)
I4
i(-2M-1,-2N-M-1)
i(1,-1) I3
y
x dummy
I2 I1 I0 source
Figure 2: a) Array of unit current sources with defined directions and positions.
b) Assignment of unit sources to specific bits for a 6-bit DAC.
Actually, it is commonly known how to select unit current sources from an array to minimize
the influence of gradient matching errors [7-8]. However, there is much less knowledge in how
the errors really effect the output signal and the overall performance.
Assuming that the matching errors are linearly distributed over the chip area [5, 6, 7] (current
source array) we can express the error sources as
δ ( a, b ) = k x ⋅ ( a – sgn ( a ) ⁄ 2 ) + k y ⋅ ( b – sgn ( b ) ⁄ 2 ) (4)
where k x and k y are constants describing the linearly graded matching properties of the unit
source array. The terms sgn ( a ) ⁄ 2 and sgn ( b ) ⁄ 2 compensate for the fact that there are no
unit sources with index a = 0 or b = 0 . We will use the unit [ LSB ] for all currents
( 1 [ LSB ] = I unit ) , and thus k x and k y will have the unit [ LSB ⁄ unit source ] . Combining Eq.
(3) and Eq. (4) we achieve an expression for ∆ k
52
∆k = ( k x ⋅ ( 2
M – 1 – 2 k + 1 + 2 k – 1 ) – ( k ⋅ ( 2 N – M – 1 – 2 – 1 ) ) ) ⋅ 2 k, 0 ≤ k ≤ M – 1
y
(5)
∆ k = – k y ⋅ ( 2 N – M – 1 – 2 k – M + 1 + 2 k – M – 1 ) ⋅ 2 k, M≤k≤N–1
From Eq. (5) we can see that the gradient in the x-direction, as could be expected, only affects
the M LSBs. Due to the 2 k factor in Eq. (5) the largest errors will occur for the MSBs, pro-
vided that k x is not very large compared to k y , which probably is a valid assumption.
0.9 0.4
0.8 0.3
0.7 0.2
0.6 0.1
DNL
INL
0.5 0
0.4 −0.1
0.3 −0.2
0.2 −0.3
0.1 −0.4
0 −0.5
50 100 150 200 250 0 50 100 150 200 250
Input word Input word
c) DNL for an 8−bit DAC with M=4, ky=0.001 and kx=0.001 d) INL for an 8−bit DAC with M=4, ky=0.001 and kx=0.001
1 0.5
0.9 0.4
0.8 0.3
0.7 0.2
0.6 0.1
DNL
0.5
INL
0
0.4
−0.1
0.3
−0.2
0.2
−0.3
0.1
−0.4
0
−0.5
50 100 150 200 250 0 50 100 150 200 250
Input word Input word
Figure 3: Typical DNL and INL plots for an 8-bit DAC with linearly graded matching errors (a and
b) only in the y-direction and (c and d) in both the x and y direction.
gradient only in the y-direction (compare Fig. 2a), with value k y = 0.001 [ LSB ⁄ unit source ]
and the width of the unit current source array is given by M = 4 . In Fig. 3a) we clearly see
that DNL is largest when turning on the MSB. The same holds for the simulation result in Fig.
3b). The INL is largest for X around 2 N – 1 , where the mismatch error also is largest (about
± ∆ N – 1 ). In Fig. 3c) and d) we show the DNL and INL when introducing a gradient error in
the x-direction as well. The gradients are given by k x = k y = 0.001 [ LSB ⁄ unit source ] . As
expected these plots are similar to those in Fig. 3a) and b). Even though the M LSBs no longer
are mutually matched, the dominating errors are still those from MSB, MSB-1 etc.
53
5. ERRORS IN THE FREQUENCY-DOMAIN
Most of the characterization of the high-speed DAC is done in the frequency-domain. Since the
distortion of a converter is dependent on input dc and amplitude level, the information on DNL
and INL is insufficient. We have to find models for the frequency-domain properties as well.
When applying a full-scale sinusoid to the DAC, the output current, I out(n) , can be written as
I out(n) ≈ 2 N – 1 ⋅ sin ( 2π ⋅ f 0 ⁄ f s ⋅ n ) + e ( n ) + I DC (7)
where f 0 is the signal frequency, f s is the sample or update frequency, I DC is the signal’s dc
level, n is the sequence index, and e ( n ) is the error signal due to the linearly graded mis-
match. Using the description of error sources from Fig. 1b) and considering all current sources,
we have that the total error signal can be written as
∑k = 0 bk(n) ⋅ ∆k
N–1
e(n) = (8)
It can be shown that
∑k = 0 ∆k ≈ –∆ N – 1
N–2
(9)
Using this information we approximate the error signal with a piecewise linear signal, ẽ(n) , as
shown in Fig. 4. Making a Fourier series expansion on the approximated error signal gives us
2
−0.5 ẽ(n)
−1 –∆ N – 1
−1.5
−2
0 1 2 3 4 5 6
Figure 4: Input signal (dashed) and approximative error signal (solid) plotted in different scales
∞ f0
∑k = 1 Ak sin 2πk -----n with A k = ∆ N – 1 ------ 1 – ------ for odd k and 0 else. (10)
4 2
ẽ ( n ) =
fs πk πk
Besides A 1 , the largest coefficient A k is given by the 3rd harmonic, A 3 , and because we have
applied a full-scale sinusoid at the input, the signal amplitude is approximately
A sig ≈ 2 N – 1 (11)
Combining Eq. (10) and Eq. (11) gives us the spurious-free dynamic range (SFDR) as
2
A signal ⁄2 9π 2 2 2 2 ( N – 1 )
-----------------------
SFDR ≈ ----------------------- = ⋅ ------------------
- (12)
A 32 ⁄ 2 4 ( 3π – 2 ) ∆ N2 – 1
Using Eq. (5) for ∆ N – 1 in Eq. (12) we get
9π 2 2
2 N – 1 ⋅ 2 M + 1 2 –2 9π 2 2 2 2M –2
SFDR = ---------------------------- ⋅ -------------------------------
- ⋅ k = --------------
- ⋅ --------- ⋅ k (13)
4 ⋅ ( 3π – 2 ) 2 2 ( N – 1 ) y 3π – 2 2 2N y
54
Rewriting Eq. (13) in dB yields
SFDR ≈ 21.6 + 6 ( M – N ) – 20 log k y (14)
The dependence on M and N has been verified in simulations. In Fig. 5a) the theoretical value
a) b)
SFDR as a function ky, kx=0 SNDR as a function of ky, kx=0
150 90
80
70
100 60
SNDR [dB]
SFDR [dB]
50 61
60.95
40
65.5 60.9
50 65 30 60.85
60.8
64.5 20 10
−4
64
−4 10
10
0 −8 −7 −6 −5 −4 −3 −2
0 −7 −6 −5 −4 −3 −2
10 10 10 10 10 10 10 10 10 10 10 10 10
Gradient in the y−direction [LSB/(unit source)] Gradient in the y−direction [LSB/(unit source)]
Figure 5: Theoretical (solid) and simulated (dashed) values of (a) SFDR and (b) SNDR as a function
of ky for a 14-bit DAC with M=8.
of SFDR is plotted as a function of k y together with a simulated value for a 14-bit DAC with
an array width of M = 8 . We see a very good agreement between theoretical and simulated
values for larger gradients, for smaller gradients the simulated values are limited by the quanti-
zation noise, which is dependent on the FFT length used in the simulations.
In order to calculate the signal-to-noise-and-distortion ratio (SNDR) we need to find the error
signal power. We could use the same approximation of the error signal as before and sum the
squares of the Fourier coefficients, but examining Eq. (10) we see that this is not easily done,
so a statistical approach is preferred. Using Eq. (8) and defining e k(n) = b k(n) ⋅ ∆ k , we get
When applying a full-scale sinusoid, the dominating covariance term, appearing twice in the
sum above, is
Cov ( e N – 1(n), e N – 2(n) ) = – ∆ N2 – 1 ⁄ 48 (16)
With a few minor approximations we also find
55
32 + 140 ⋅ 4 M – N 1 2 2 8
P ND = ------ + ∆ N – 1 ⋅ ----------------------------------------- – ------ – 1 – --- ⋅ -----2
1 2
12 105 24 π π (20)
Using Eq. (5) and Eq. (11) we now get
2 ⁄2
A sig 2 2N – 3
SNDR = ----------------- = -----------------------------------------------------------------------------------------------------------------------------------------------
- (21)
P ND 1 2 4 ( N – 1 ) 32 + 140 ⋅ 4 M – N 1 2 2 8
------ + k y ⋅ -------------------
2 - ⋅ ----------------------------------------- – ------ – 1 – --- ⋅ -----2
12 22( M + 1) 105 24 π π
In Fig. 5b) SNDR is plotted in the same way as for the SFDR. We see a very good agreement
between the theoretical expression and the simulated values. At small gradients, the SNDR is
determined by the signal-to-quantization noise ratio SQNR ≈ 6 ⋅ 14 + 1.76 ≈ 86 dB. We also
see that at larger gradients when M – N ≥ 3 , Eq. (21) may be approximated by
SNDR ≈ 17 + 6 ( M – N ) – 20 log k y (22)
6. CONCLUSIONS
We have derived formulas describing the frequency-domain performance of current-steering
CMOS DAC when applying linearly graded matching errors to unit current source elements.
Calculated and simulated results correlate very well. To decrease the influence of linearly
graded matching properties, different techniques and layout styles have been proposed (e.g.
[7]). However, it is desirable to make the design as simple as possible, and given expected
parameter variations one can predict what type of technique or layout style that is needed to
reach a certain goal in terms of DAC performance. The DAC structure analyzed in this paper is
rather naive. For example the lack of segmentation may increase glitches, but the approach
used to treat deterministic mismatch could also be applied to other structures and combined
with models of other error sources such as random mismatch, wire resistance, and finite output
impedance we are able to find a very good model of current-steering DACs, a model that can
be used to design high performance DACs without too much trial and error.
The authors would like to thank Niklas U. Andersson and Pierre Dalheim-Lander at
Linköpings universitet, Sweden, for their help.
7. REFERENCES
[1] P. Hendriks, “Specifying communication DACs,” IEEE Spectrum, Vol. 34, No. 7, pp. 58-
69, July 1997
[2] D.A. Johns and K. Martin, Analog integrated circuit design, John Wiley & Sons, New
York, NY, USA, 1997
[3] R.J. van de Plassche, Integrated analog-to-digital and digital-to-analog converters, Klu-
wer Academic Publishers, Boston, MA, USA, 1994
[4] J.J. Wikner and N. Tan, “Modeling of CMOS D/A Converters for Telecommunication,”
IEEE T. of Circuits and Systems II, May 1999
[5] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, “Matching Properties of
MOS Transistors,” IEEE J. of Solid-State Circuits, vol. 24, pp. 1433-39, Oct. 1989.
[6] J. Bastos, et al, “Mismatch Characterization of Submicron MOS Transistors,” Analog
Integrated Circuits and Signal Processing, Vol. 12, pp. 95-106, 1997
[7] H. J. Schouwenaars, et al, “A Low-Power Stereo 16-bit CMOS D/A Converter for Digi-
tal Audio,” IEEE J. of Solid-State Circuits, Vol. 23, pp. 1290-1297, Dec. 1988.
[8] T. Miki et al, “An 80 MHz 8-bit CMOS D/A Converter,” IEEE J. of Solid-State Circuits,
Vol. 21, pp. 983-988, Dec. 1986.
56
References
[1] T. Starr, J.M.Cioffi, and P.J. Silverman, Understanding Digital Subscriber Line
Technology, Prentice Hall PTR, Upper Saddle River, NJ, USA, 1999, ISBN 0-13-
780545-4
[2] D.A. Johns and K.Martin, Analog Integrated Circuit Design, John Wiley & Sons, New
York, NY, USA, 1997, ISBN 0-471-14448-7
[3] P.Hendriks, “Specifying Communications DACs”, IEEE Spectrum, vol. 34, no. 7, pp.
58-69, July 1997.
[6] R.J. van de Plassche, Integrated analog-to-digital and digital-to-analog converters, Klu-
wer Academic Publishers, Boston, MA, USA, 1994
[7] K. Lin, and C.H. Bult, “A 10b 250MSample/s CMOS DAC in 1mm2”, in Proc. of the
1998 IEEE Int’l Solid-State Circuits Conf. (ISSCC’98), pp.214-15, San Fransisco, CA,
USA, Feb 1998
[8] B.G. Streetman, Solid State Electronic Devices, Prentice-Hall Inc, Englewood Cliffs,
NJ, USA, 1995, ISBN 0-13-436379-5
[9] M.J.M. Pelgrom, A.C.J. Duinmaijer, and A.P.G. Welbers, “Matching Properties of
MOS Transistors”, IEEE J. of Solid-State Circuits, vol. 24, no. 5, pp. 1433-9, Oct 1989
[10] C.A.A. Bastiaansen, D.W.J. Groeneveld, H.J. Schouwenaars, and H.A.H. Termeer
“A 10-b 40-MHz 0.8 µ m CMOS Current-Output D/A Converter”, IEEE J. of Solid-
State Circuits, vol. 26, no. 7, pp. 917-21, Jul 1991
[12] H.J. Schouwenaars, D.W.J. Groeneveld, and H.A.H. Termeer “A Low-Power Stereo
16-bit CMOS D/A Converter for Digital Audio”, IEEE J. of Solid-State Circuits, vol.
23, no. 6, pp. 1290-7, Dec 1988
[16] P. Dahlheim Lander, Design of an MSB Randomizer for Wide Band CMOS DAC,
Final Work, Linköping University, Sweden, 1998, LiTH-ISY-EX-ET-0159
[17] S.M. Kang, and Y. Leblebici, CMOS Digital Integrated Circuits: Analysis and Design,
McGraw-Hill, Singapore, 1996, ISBN 0-07-114423-4
[18] N.H.E. Weste, and K. Eshraghian, Principles of CMOS VLSI Design: A Systems Per-
spective, Addison-Wesley, USA, 1994, ISBN 0-201-53376-6
Design of 12-bit 100-MHz Current-Steering DAC for SOC Applications
Chun-Yueh Huang Tsung-Tien Hou, and Chi-Chieh Chuang Hung-Yu Wang
Department of Electronic Engineering Chip Implementation Center,
Kun Shan Universiv of Technology National Applied Research
Yung-Kang, Tainan, 71003 Taiwan, R.O.C. Laboratories,
cyhun@,mail.ksut.edu.tw Taiwun, R.0.C.
ted@,tn.cic.edu.tw
117
0-7695-2403-6/05$20.00 0 2005 IEEE
For ease of understanding the design methodology,
the basic components of the current-steering DAC are
discussed in detail as follows.
0
CLK
$j-M2
Do 1
1
NOT3 NOT4
A. D-latch
The circuit diagram of D-latch is shown in Fig. 2. Figure 3. The clock tree structure
The D-Latch, which is the simplest digital storage
element, is a samplehold circuit and can be used to I". .+. ................................ i
keep the synchronizationof control signals in the DAC.
Basically, in the D-Latch, two cross-coupled inverters thermometer s1, sz, .... s,".,
(NOTl and NOT2) are acted as a storage unit and two encoder
inverters (NOT3 and NOT4) as buffers. When CLK is . s,"
high, transistors M 2 and M3 are turned on, and the thermometer s,"+,,s,",,, .... S F ' 4
input signal D through transistors M 2 and M3 are encoder
... ............................. ................... I..........
."
118
......
. n.nnmn
D. Current cell In Fig. 5 , the output impedance of the current cell
Routis approximately calculated as
In the DAC, the current source cannot be Rout gm4gm3gm2ro4ro3ro2rol (3)
completely turned off and therefore the switching
signals have to be properly matched to improve the In the current cell, the device dimensions of the
glitch performance. A proper switching scheme for the channel width W and the channel length L for M1-M2
PMOS differential current switch is implemented by and M3-M6 are 6.Y2.5 and 3/0.35 pm, respectively.
transistors M3 and M5 as shown in Fig.5. Two Next, when we implement the DAC layout, we have
transistors M4 and M6 with self-bias are used to to think of some questions. For instance, the current
increase the output impedance of the current cell. cell will exist transistor matching errors, i.e., size
The output impedance of the current cell has a great errors, threshold voltage variations, supply and bias
influence on the specifications of INL (integral non- voltage variations, oxide thickness variations, output
linearity) and SFDR [9,10]. The relationship between voltage variations, etc. In order to improve transistor
the impedance and the achievable INL specification is matching errors, in the current source array, we use
given by [ 101 basic unit current sources from different positions to
T R AT2 construct this array by means of common-centroid
technique as shown in Fig. 6. This approach can
reduce the problem of process graded and systematic
where RL is the load resistor, runitis the LSB current, errors, but it will increase the wire resistance and
and N is the total number of unit current sources. capacitance as well. The shaded area is the dummy
Considering the frequency dependency of the cells for avoiding over etching of process.
output impedance, the relation between the SFDR
specificationis given by [9]
R,, =-NRL
4s
where S is the ratio between the fundamental signal
and the second harmonic component caused by the
output impedance effect. According to Eq. (2), for a
12-bit current steering DAC, the ratio has to be at least
equal to 72 dB. If the load resistor is a 50!2 double
terminated cable and N is equal to 4095, the value for
required output impedance Routhas to be at least lOOM
!2 in the Nyquist frequency range. In order to make Figure 6 . The floor plan common-centroid technique
the output impedance of the current cell large enough, of current source
and to satisfy the specification of INL and SFDR, the
cascade current source configuration and the cascade E. Bias circuit
switching transistors as shown in Fig. 5 are used to
design the current cell. In Fig.5, the cascade current mirror in the current
cell can be used to reduce short-channel effects and
increase the output impedance, but it will limit the
signal swings. In order to reduce this limitation, a
wide-swing cascade current mirror bias scheme as
shown in Fig. 7 is used to provide the accurate bias
voltages for the current cells. In the bias circuit, the
diode-connected transistors M1, M2, M3 and M8, M9,
M10 provide a bias to transistor M4 and M12,
respectively. The transistors M4 and M12 are used to
increase the gate-source voltages of transistor M5 and
M11 and control the drain-source voltages of transistor
Figure 5. The circuit diagram of a current cell M5 and M11 in edge of saturation.
119
I ..
The above approach can reduce the variations in the 0.25 LSB and 0.4 LSB, respectively, as shown in Fig.
drain-source voltage of M5 and M7 with changes in 9 andFig. 10.
VDD,and hence can make the bias current rout more For the SFDR simulation, the converter is simulated
accurately equal to the reference current Im~. The at the condition of the update rate for a full-scale input
reference current IREF depends on external resistor R signal with a fkequency of 1 MHz sine wave. Fig. 11
and the device dimensions of the transistor M5.The shows that the SFDR remains above 80 dB at the
value of IREF can be derived as follows update rate of 100 MS/s. Furthermore, Fig. 12 shows
‘DLI = ‘REF + G‘ S5 (4) that the SFDR is above 65dE3, for Nyquist fi-equency
range simulation, in the condition of a full-scale input
signal with a frequency of 49 MHz under the sampling
rate of 100 MHz.
The settling time of the DAC is less than 911s.The
power consumption is 47 mW at the maximum
where K = O.S,uu,C,W I L , V,, and VT are the conversion rate. The summary of the performance of
transconductance parameter, the gate-to-source voltage, the proposed DAC is shown in Table 1.
and the threshold voltage, respectively. Then, the gate-
to-source voltage of V,,, can be rewritten as
M1
M2
M3
3. Simulation results
The layout of the proposed 12-bit lOOMHz current-
steering DAC has been finished as shown in Fig. 8.
This circuit is simulated by HSPICE based on TSMC
standard 0.35um 2p4m CMOS process [ll]. The
power supply is 3.3 Volts and the output load resistor
is 50 SZ. The current of LSB is designed as 3 pA.The
post layout simulation results show that the static
performances of DAC in DNL and INL are less than
Figure 9. Post layout simulation for DNL of DAC
120
. . -,
. , I I..
Table 1. Summary of the DAC performance
, . * .
4. Conclusion
. , " ~ .
~ .^ In this paper, a 12-bit 100-MHz current-mode
digital-to-analog converter (DAC) for system-on-a-
chip (SOC) applications is presented. The SFDR is
larger than 65dB for a Nyquist fimdamental signal.
Based on the TSMC 0.35um 2p4m CMOS technology,
the proposed DAC is simulated by HSPICE. The post
layout simulation results show that the proposed DAC
has the following characteristics: INL,< kO.4LSB,
DNL<3.25LSB, and settling time less than 9ns. The
power consumption is 47mW at the maximum rate.
The chip area is 6.9mm2. In the near hture, the chip
Figure 1 1. The SFDR for a 1MHz signal at a will be fabricated. After finishing the chip fabrication
1OOMHz update rate. and testing the chip's function, we will develop the
corresponding Verilog-A model of DAC in SOC
applications.
.. .. . . .. . ^. . ... . .* .
" *
5. References
[l] J. Vandenbussche, G . Van der Plas, et al., "Systematic
Design of High-Accuracy Current-Stecring D/A Converter
Macrocells for integrated VLSI System", IEEE Trans.
Circuit and Syst. II, vol. 48 NO. 3, March 2001.
121
[4] B. J. Tesch and J. C. Garcia, “A Low Glitch 14-b [8] A. Van den Bosch et al., “A 10-bit 1-Gsamplesls nyquist
lOOMHz DIA Converter,” IEEE J. Solid-state Circuit, Vol. current-steering CMOS DIA converter,” IEEE J. Solid-state
32, pp. 249-252. Circuits, ~01.36,pp. 3 15-324, Mar. 2001.
[5] M. Gustavsson, J. J. Wilner and N. N. Tan, CMOS Data [9] A. Van den Bosch, M. Steyaert and W. Sansen “SFDR-
Converters For Communications. Boston, Kluwer Academic bandwidth limitations for high speed high resolution current
Publishers, 2000. steering cmos d/a converter,” in Proc. IEEE Int. Con$
Electronics, Circuit and System (TCECS), Sept. 1999, pp.
[6] C. H. Lin and K. Bult, “A lob 250Msmapleh CMOS 1193-1196.
DAC in lmm2,” in Proc. 1998 Int. Solid-state Circuit Con$
(ISSCC),pp. 214-215, Feb. 1998. [lo] B. Razavi, Principles of Data Conversion System
Design, IEEE Press, 1995. ISBN 0-7803-1093-4.
[7] C. H. Lin and K. Bult “A 10-b, 500-Msampleh CMOS
DAC in 0.6 m 2 , ” IEEE J. Solid-state Circuits, Vol. 33, No. [l11 R. Gregorian, An Introduction to Mixed- ,Signal IC Test
12, Dec. 1998. and Measurement. New York, Oxford, 2001.
ISCAS 2000 - IEEE International Symposium on Circuits and Systems, May 28-31, 2000, Geneva, Switzerland
0-7803-5482-6/99/$10.0002000 E E E
11-745
digitallselection logic to be able to automate the Fig.3 : the quadratic standard deviation of the current
measurements to a large extent [2,3]. The dimensions of factor in function of the inverse of the square root of the
area of the measured transistors
VTniUraM A*c-725>mv”m
11-746
schematic for one unit current cell is depicted in fig.4. The generator (HP80000)so that for different values for the
design parameters of the D/A converter will be calculated (VGs-VT)csthe current at the output of the D/A converter
for a yield estimation of 99.7 % and an output voltage will be measured (HP3457 multimeter) and
swing of 1 Volt peak to peak into a 50 R load. This leads
to a full scale drain-source current specification of 20 mb.
(=IFS).
(3)
I line fined through the
measurement points /
Since the transistor matching parameters seem to improve
for smaller gate-length technologies, the parameters of the
0.7 pm technology can be used in a first approximation for
the design of a 0.5 pm D/A converter. Apart from the
mismatch parameters, the gate overdrive voltage of the
current source transistors has to be determined. An
l/(vGs-<T)z
appropriate value can be found by making a trade-off
between the settling time of the D/A converter and the chip Fig.5 : the quadratic relative standard deviation in
area [5]. The optimal value for the gate-source voltage of function of the inverse quadratic gate overdrive voltage
the current source is then given by a value of of the current source transistors
approximately 1.6V. If the operation frequency of the D/A
converter is of minor importance, the value of the From the IDs-V~s characteristic, the formula for the relative
overdrive voltage can be increased leading to the use of a current standard deviation is easily derived :
smaller silicon area.
The values for the dimensions of the current sources
(4)
(implemented in the 0.5 pm technology) for both the
optimal gate overdrive voltage and for a larger value can For large values of the gate overdrive voltage this formula
be found in table 1. simplifies to :
11-747
On the other hand, for small values of ( V G S - V ~the )~~ These results are in good agreement with the results
influence of the current factor is negligible and the relative obtained from the test structures described in paragraph 2.
unit current standard deviation is given by : This proves that the D/A converter can indeed be
successfully used as a process monitor for mismatch.
4. CONCLUSION
Since the characteristic of fig.3 is a straight line, a fitting of In this paper the CMOS current-steering D/A converter has
d(AI/I) calculated from the measurement data gives us the been presented as a test structure for matching purposes.
threshold voltage mismatch coefficient. Using a D/A converter as a test structure has a large cost
advantage since no especially designed test structures will
be necessary anymore. The time to implement the.se
circuits and the silicon area they consume can now be used
to design a chip (=DAC) that can be further used in e.g.
telecommunication applications.
The distance term in Pelgrom’s equation can be neglected The extraction procedure of the transistor mismatch
because the current sources are switched following an S parameters has been described and is verified by
pattern. The current source that in the layout is placed comparing the measurement results of a six bit current-
beside to the previous measured current source is measured steering D/A converter and the results obtained from the
next. Furthermore, to avoid the influence of edge effects classical test structures.
dummy rows and columns are used in current-steering D/A
converters. This guarantees the fact that all measured 5. ACKNOWLEDGEMENT
transistor pairs have identical surroundings.
The authors wish to thank Alcatel Microelectronics for
3.4 Measurement Results their support.
11-748
Folded-Current-Steering DAC: An Approach to Low-
Voltage High-Speed High-Resolution D/A Converters
Soheil Radiom Behzad Sheikholeslami Hamed Aminzadeh Reza Lotfi
Integrated Systems Lab., IC-Design Lab., Integrated Systems Lab., Integrated Systems Lab.,
EE Dept. ECE Dept. EE Dept. EE Dept.
Ferdowsi University University of Tehran Ferdowsi University Ferdowsi University
Mashhad, I.R.Iran Tehran, I.R.Iran Mashhad, I.R.Iran Mashhad, I.R.Iran
soheil_radiom @ieee.org
Abstract- In this paper, a new topology for CMOS current In section II, the current-steering D/A converters, their
steering DAC, i.e. folded current-steering structure, suitable basics and their conventional structure are described. Section
for low-voltage applications is presented. Using the folded III introduces the proposed folded structure. In section IV
structure, more voltage headroom will be available for the simulation results confirming the effectiveness of the
current-source devices increasing the output impedance of the proposed approach are presented following by conclusions in
current sources. The DAC linearity is thus improved. Also this section V.
novel topology improves the value of PSRR and output voltage
swing without SFDR degradation and results in more constant
load for current cell. Also a current-steering logic decoder with I VARIOUS CURRENT-STEERING TOPOLOGIES
differential outputs is suggested. Besides, gain boosting is Current-steering DACs are based on an array of the
applied to increase the output impedance of current source. matched current sources that are switched to the output node
These techniques are used to design a 12-bit 200MS/s DAC based on the input digital code. Three different ways of
with 1.5-V power supply and more than 72dB SFDR is implementation are available, namely binary, unary and
achieved in all frequencies. Also Monte-Carlo simulations show segmented architecture. Each of the architectures will be
that the probability of an SFDR of more than 65dB is more briefly discussed.
than 90% for an input frequency of 100MHz.
A. The Binary Weighted Architecture
I. INTRODUCTION In the binary implementation, the binary digital input
The recent growth in the market demand for VLSI codes directly control N binary-weighted current sources, i.e.
circuits and also the advance in implementing digital each switch conveys a current to the output node that is twice
integrated circuits have made the interface between the larger than the current of the previous least significant bit.
analog and digital parts of a system as one of the most This architecture suffers from the poor matching
challenging blocks of the system to design. Applications in performance of non-equal current sources. Large glitches are
the area of video and wireless communications need DACs also expected in this structure.
with high performance. For these applications, among
different architectures, the current-steering configuration is B. The Unary Architecture ( Thermometer coded)
frequently chosen due to its superior performance from the In the unary architecture, the current sources are switched
speed, resolution and power consumption points of view. On to the output node based on the thermometer version of the
the other hand, designing high-speed high-resolution data input binary code. 2N -1 unit current sources, equal to ILSB
converters is more challenging with lower supply voltages, are switched. The matching performance is much better than
Although at low frequencies the output impedance and the binary architecture and the glitches are smaller and code
mismatch between current sources have the main influence dependent; however the decoder complexity is added to the
on linearity [1], in high-speed, high- resolution DACs, the circuit.
most important parameter for evaluating the performance is
Spurious Free Dynamic Range (SFDR) which is mostly C. Segmented Architecture
limited by finite output impedance, charge feedthrough,e th
major carry glitches and spikes in common switch nodes and In order to employ the advantages of both architectures
power supply noise. the segmented topology is proposed, i.e. the binary
architecture is employed for less significant bits whereas the
B6 B7 l i
Thermometer Decoder
_ £ £
||.
4784
current ofthe load resistors is again proportional to the input Vdd
digital code. However, there are several advantages.
As the drain-source voltage of the fold transistors can be
adjusted, the headroom voltage on the switches and current
sources is increased, i.e. larger VDS for CS, CAS and switch
transistors are provided, so higher SFDR and more linearity
as a result of increase in current source output impedance is
obtained. Also, this increase in their VDS leads to better Gate of Mca
PSRR. On the other hand, increasing the output impedance Source of Mcas
will improve the performance of the current sources as well
and make their behavior more robust against the data-
dependent transitions caused by input signal variations on
node V, [5]. Moreover what transistor Mk does is to shield 'r
the critical node (Y) from loading effect of Rload and the
effect of the output voltage swing on the performance of
current sources and thus to provide a more constant load for Fig.4. The applied gain booster for increasing output impedance
current cell. Also it reduces the effect of input code spikes on
the output nodes. VDD
Here the effect of charge feedthrough which seriously
degrades the performance of conventional DACs [6] is
considerably decreased in the output node as there is no CLK CLK
direct path to the output. So there is no need for the cascode VIP T_'_L__
transistor above switches that was offered in [4] which limits
the swing. Also here one can get more swing without SFDR
degradation as a consequence of any limitation on VDS of
current cell transistors.
The CMFB used in the folded branch is for compensating
the variations in various temperatures.
Another technique applied to the circuit is gain boosting
which effectively increases the output impedance and leads Fig.5. Latch circuit
to more SFDR and better linear behavior. For obtaining more
gain the boosting amplifier of Fig. 4 is used in the circuit. Another proposed technique employed in the circuit is
using current-steering logic with differential outputs for
binary to thermometer decoder that significantly reduces the
effect of the decoder output voltage spikes in the circuit.
VDD
Faster transitions push the spurious outputs to out-of-
band frequencies. Besides, fast switching can decrease the
transition energy. The charge injected by the capacitive
FVCMFB + coupling ofthe digital signals to the output is independent of
1.I /1 I_ MFOLD the switching speed, but the time ofunbalanced switching is
reduced [3] and this is the reason of choosing NMOS
0 * 0 0 yswitches here. Considering the fact that in the offered
r -./-11-/ 1MK structure the direct path between input signal and output is
F,,
> rl removed, the effect of glitch energy in the output node
IV\
x I @ _ substantially decreases.
dump
Lg out The circuit depicted in Fig.5 has been employed for the
CAS 41-0 T tlatch circuit, previously proposed in [7].
4785
swing is 1.2 Vp-p,djf with a 25-Q load and a power-supply V. CONCLUSION
voltage of 1.5V. At 200 MSample/s the analog and digital In this paper, a new topology for current steering DACs
parts consume 85mW. Monte-Carlo simulations considering was presented. This topology results in larger VDS on
the mismatches of the current sources show that the current-source transistors which causes higher output
probability of an SFDR of more than 65dB for a Nyquist-
frqecipti mor tha 900 frequenc input is more impedance and as a consequence better performance and
than90PSRR. Also this method provides constant load for current
cells and improves output voltage swing without SFDR
Resolution 12 degradation. Also the proposed techniques for the binary to
thermometer decoder, using current steering logic with
Decoding Segmented (6+6) differential outputs, as well as gain-boosted current sources
Nyquist update frequency 200MS/s improve the circuit behavior considerably. Simulation results
Differential output signal 1.2Vpp of a 12-bit 200MHz DAC confirm the effectiveness of the
Supply voltage 1.5V proposed techniques.
Iload 25mA
SFDR up to Nyquist freq. > 72 dB
INL 1LSB ACKNOWLEDGMENT
DNL 0.7 LSB The authors would like to express their appreciation for
Total power consumption @ 85 mW Dr. Robert Neff and Dr.Ola Anderson for providing ongoing
Nyquist rate guidance.
Technology 0.18prm CMOS
77 REFERENCES
[1] K.Doris, J.Briaire, D. leenaerts, M.Vertregt, A.van Roermund "A 12b
76 - 500MS/s DAC WITH >70dB SFDR up to 120 MHz in 0.18um
CMOS," IEEE International Solid-State Circuit Conference 2005.
75 - \ [2] Chi-Hung Lin and Klaas Bult " A 10-b 500-MSample/s CMOS DAC
in 0.6mm2," IEEE JOURNAL OF SOLID-STATE
74 - \ CIRCUITS,VOL.33, NO 12 ,DECEMBER 1998.
[3] Jurgen Deveugele, Michiel Steyaert " A 10b 250 MS/s Binary-
Weighted Current-Steering DAC," IEEE International Solid-State
a_ \; / , _+,Circuit Conference 2004.
72 - [4] A.Van den Bosch, M.Steyaert and W.Sansen "SFDR-Bandwidth
Limitation for high speed high resolution Current Steering CMOS
71 - D/A Converter,"
[5] W.Schofield "A 16b 400MS/s DAC with <-8OdBc IMD to 300MHz
70 - and <-160dBm/Hz Noise Spectral Density," ISSCC Dig.Tech.papers,
Feb.2003.
69 I_ I_ I_ I_ I_ I_I [6] Van Beek, K.Doris, J.A.Hegt and Van Roermund "Optimum
3/64 7/64 11/64 15/64 19/64 25/64 29/64 31/64 Segmentation for high speed current steering Digital-to-Analog
Converters,"
f/fs
[7] A.Van den Bosch, F.Borremans, Michel Steyaert , Willy Sensen " A
10-bit 1-GSample/s Nyquist Current-Steering CMOS D/A
Converter," IEEE JOURNAL OF SOLID-STATE
Fig.6 SFDR VS signal frequency. The sample rate is 200MS/s CIRCUITS,VOL36,No.3, MARCH 2001.
4786
Linköping Studies in Science and Technology
Thesis No. 715
J Jacob Wikner
LiU-Tek-Lic-1998:50
J Jacob Wikner
LiU-Tek-Lic-1998:50
The general basics of digital-to-analog converters are described in this thesis. A combi-
nation of behavioral-level, circuit-level, and layout-level descriptions, as well as design
methodologies are presented. A thorough description of the design of a CMOS Nyquist-
rate digital-to-analog converter chipset for wideband radio and VDSL applications, and
of a CMOS oversampling digital-to-analog converter for ADSL applications is given.
Measurement results are presented as well as suggestions on how to improve perfor-
mance.
In these kinds of applications, it is the dynamic performance of the converters that deter-
mines the quality of the device. Commonly, the characterization of the converters is done
in the frequency domain. We have focused on how static and dynamic properties are
affecting the output spectrum. In the thesis we present important work on how these
effects are modeled and how they apply to a CMOS current-steering digital-to-analog
converter structure as used in the Nyquist-rate chipset. It is shown how the output imped-
ance, matching, and noise are affecting the performance when changing fundamental
component values of the circuit. This work has resulted in four publications, and the
results can be used as a guide when designing high-resolution converters.
Since large mixed analog/digital circuits contain both complex analog and digital parts
it is important to be able to simulate and vary parameters on a higher behavioral-level.
This makes it possible to use faster simulation tools to find structures near the optimum.
With time-consuming circuit-level simulations these structures are only verified. This
allows a faster design flow. Especially during the design of the oversampling converter
this design strategy has been used and MatLab models have been developed.
Part A: Acknowledgment
I would like to thank my supervisor, Ph. D. Nianxiong Tan, for all the great help and
encouragement he has given me. As he said, “... before you know how to drive a car, you
first have to crash it some times ...”
The help with discussions and to make the working day more interesting, the whole
group, Electronics Systems at Linköping University, with our supervisor Prof. Ph. D.
Lars Wanhammar, is acknowledged. I would especially like to thank Mikael Gustavsson
for valuable discussions and great help. All the help from former Ph. D. students at
Electronics Systems (former Applied Electronics) is acknowledged.
Very much thank you, the Ericsson people; Gunnar Björklund and Jan-Erik Eklund
(MERC, Ericsson Components AB), and Bengt E. Jonsson (Ericsson Radio AB). Thanks
for the great help with Cadence; Helge Stenström (Ericsson Radio AB), and for the guide
into the mysterious world of measurements; Peter Pettersson (Ericsson Radio AB).
The work is financially supported by the Swedish National Board for Industrial and
Technical Development (NUTEK) and the Strategic Research Fund (SSF).
CMOS
Digital-to-Analog Converters for
Telecommunication Applications
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Outline of the Thesis ................................................................................ 1
1.2 Papers and Conference Contributions ...................................................... 2
1.3 Nomenclature, Acronyms, and Abbreviations ......................................... 3
Part A:
Overview:
Digital-to-Analog Converter Design
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1 Research Background .............................................................................. 5
1.1.1 The DAC 5
1.1.2 The Telecommunication Situation 6
1.1.3 Higher-Level Design 8
1.2 Signals and Coding Schemes ................................................................... 9
1.2.1 Signal Definitions 9
1.2.2 Digital Numbers and Representation 10
1.3 xDSL Standards ..................................................................................... 12
1.4 The CMOS Transistor ............................................................................ 14
1.4.1 Operation Regions and Small Signal Characteristics 15
i
Table of Contents
Part B:
Nyquist-Rate Digital-to-Analog
Converters for VDSL Applications
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2 Nyquist-Rate DAC Structures . . . . . . . . . . . . . . . . . . . . . . . 39
2.1 Charge-Redistribution DACs ................................................................ 39
2.2 Current-Steering DACs ......................................................................... 40
2.3 R-2R Ladder DACs ............................................................................... 41
ii
Table of Contents
Part C:
Oversampling Digital-to-Analog
Converters for ADSL Applications
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
2 Oversampling DAC
Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.1 Interpolation Filters ................................................................................ 84
2.2 Modulator ............................................................................................... 86
iii
Table of Contents
Conclusions
References
iv
CMOS
Digital-to-Analog
Converters for
Telecommunication
Applications
1 Introduction
This thesis is divided into five parts and is outlined in Sec. 1.1. Except this introduction
and the conclusions’ part, the thesis consists of three major parts as described below.
1
CMOS Digital-to-Analog Converters for Telecommunication Applications
2
Introduction
Notation Description
F{ } Fourier transform
L{ } Laplace transform
R Set of real numbers
t≥0
u(t ) Step function, u ( t ) = 1
0 t<0
Z Set of integer numbers
β Transconductance parameter, β = K' ⋅ ( W ⁄ L )
C ox Oxide capacitance per gate area
fN Nyquist frequency, f N = f s ⁄ 2
f OSR Oversampling frequency, f OSR = OSR ⋅ f s
fs Sampling or update frequency
gm Small-signal transconductance
3
CMOS Digital-to-Analog Converters for Telecommunication Applications
Notation Description
HDSL High data rate digital subscriber line
HDTV High definition television
HP Highpass (filter)
IIR Infinite-length impulse response
IFFT Inverse fast Fourier transformation
IMD Intermodulation distortion
INL Integral non-linearity
LP Lowpass (filter)
LSB Least significant bit
MSB Most significant bit
NTF Noise transfer Function
OFDM Orthogonal frequency division multiplexing
OSR Oversampling ratio
PAM Pulse amplitude modulation
PCM Pulse code modulation
PSD Power spectral density
THD Total harmonic distortion
SC Switched capacitor
SI Switched current
SFDR Spurious-free dynamic range
SNDR Signal-to-noise-and-distortion ratio
SNR Signal-to-noise ratio
STF Signal transfer function
VDSL Very high data rate digital subscriber line
Table 1-1: Constants, values and variables used in the thesis.
4
Overview:
A
Part A:
Digital-to-Analog
Converter Design
1 Introduction
In this introductory part we describe and motivate the research background to the mate-
rial presented in the thesis. An overall view of the telecommunication situation is given,
and especially the roles that the digital-to-analog converter (DAC) and analog-to-digital
converter (ADC) play in the front-end and back-end of a transceiver are discussed. In
some way, the digitally processed information has to be transmitted over a channel using
only analog signals and a certain transmission technique. Therefore, the use of ADCs and
DACs is unavoidable in telecommunication applications, and the converters may very
well be the bottleneck of the whole system.
The research background, digital codes, xDSL standards, the CMOS transistor, and cir-
cuit noise are discussed in this chapter.
5
Part A: Overview: Digital-to-Analog Converter Design
For a single ended output Nyquist-rate converter with an input word length of N bits, the
amplitude level at the output for an offset binary code would be
where A ref is the reference level for the least significant bit (LSB), x 0 , i.e., the amplitude
value at the output when only the LSB is active, and x N – 1 is the most significant bit
(MSB) in the digital input word X = ( x N – 1, x N – 1, …, x 1, x 0 ) . Naturally, there are sev-
eral ways to implement a converter of this kind [23, 37, 56, 58]. This is further discussed
in Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications.
The converter may also be constructed as a finite-state machine where the output
depends on previous states of the converter. This is used in feedback strategies or special
assigned converters as for example an oversampling digital-to-analog converter
(OSDAC) that uses pulse coded modulation (PCM) [11, 56]. It is further described in
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications.
Future research will focus on how the signal information can be used to improve the per-
formance of the converter. With feedback strategies the output of the circuit is sensed and
with that information we vary parameters controlling the transfer function of the converter
and the performance may be improved. A use of redundancy can also improve the results.
By extending the digital code into more bits and use a higher resolution we can use redun-
dant parts of the converter to correct errors (non-linearities) that are generated [56]. Cal-
ibration techniques have been implemented in audio DACs which have a smaller
bandwidth [11]. For wideband DACs where the bandwidth is large there might be prob-
lems with using feedback since it limits the bandwidth and reduces speed. Some transmis-
sion schemes use time-slots during which no information is sent. For some calibration
techniques it may be possible to use samples of the generated output signal at a slower
rate than the transmission rate, i.e., using the time-slots where no information is sent. The
problem still, is that the output of a DAC is analog and that the correction most often is
done digitally. In an implementation, extra analog-to-digital conversion that might be
needed for a specific correction technique should be avoided, since this may increase the
complexity dramatically.
6
.
ers, filters, and amplifiers, which have to interact with each other. In the transmitter, the
digital code vectors are transformed into a proper analog representation and then put on
the channel, which finally is reversely transformed by the receiver.
Channel
Figure 1-2: Classical view of communication system. The system can be divided into digital parts,
mixed analog/digital parts, and analog parts.
Mixed analog/digital circuits also introduce some additional problems [33], e.g., inter-
ference from the digital circuits on the analog circuits, noise from supply lines, etc. Spe-
cial care has to be taken during design and simulation of these circuits.
Another approach is the so called coded modulation, where the source, encoder, and
modulator are no longer optimized separately. Instead the ambition is to optimize the
whole transmitter (and receiver) as one unit. This also implies new possibilities for the
design of mixed analog/digital parts, where new simulation and design methodologies
may be used.
In this thesis we focus on the digital subscriber line standards [83-88], which commonly
are referred to as xDSL which is the overall notation for several standards (or de facto
standards) as ADSL, HDSL, VDSL, etc. The general structure of a transceiver for wide-
band radio and very high data rate digital subscriber line (VDSL) is sketched in Fig. 1-3
and a block view of a modem for asymmetric digital subscriber line (ADSL) is shown in
Fig. 1-4. Generally, the existing copper wire telephone nets are to be used for the xDSL
applications, as is further described in Sec. 1.3.
Traditionally, video DACs have been used in telecommunication applications, since the
transmission speed for earlier applications has been low enough [17, 22, 52]. As the trans-
mission speed increases as well as the number of users, the requirements on the converters
are much higher. The converters are specially designed, assigned for this specific situa-
7
Part A: Overview: Digital-to-Analog Converter Design
tion. In the future we will probably see more specific circuits optimizing the whole system
with the converters, amplifiers, and filters at the front-end and back-end of the transceiver,
with the goal to minimize a cost function as for example the overall power consumption.
The ambition is also to be able to put the DAC and/or ADC directly on the line or antenna
to minimize the number of analog components, such as mixers, modulators, and amplifi-
ers [15].
Splitter
Digital
Subscriber
Pots
Line
Interface
Wideband Analog
AGC
A/D Filter
Pots Interface
Mixed Analog/Digital System
Figure 1-4: Modem for ADSL applications over the existing telephone system.
8
.
Simulation &
Specifications Verification
Layout-Level
Implementation
Behavioral-Level
Description & “Faster”
Simulation Design
Circuit-Level
Flow
Circuit-Level Implementation
Models
Layout-Level
Models
Figure 1-5: Behavioral-level design and simulation provides a faster design flow.
a(t) T
a(n)
a) t b) n
With an analog signal we understand a signal that is both continuous-time and continu-
ous-amplitude. The analog signal is described by a mapping function, f , from the time
point, t , to the amplitude A = f ( t ) as
t → ( A = f (t)) (1-2)
nT → ( A = f ( nT ) ) (1-3)
where n ∈ Z and A ∈ A L , Z is the set of integer numbers, A L is the set of possible am-
plitude levels determined by the digital code used, and T is the time period between two
9
Part A: Overview: Digital-to-Analog Converter Design
consecutive values.
X = ( x N – 1, x N – 2, …, x 2, x 1, x 0 ) (1-4)
where x N – 1 is the MSB and x 0 is the LSB. For signed digital numbers we refer to x N – 1
as the sign bit. We only discuss the integer numbers in this thesis.
In an offset binary code a number is given by
N–1
X = ∑ 2k ⋅ xk = x 0 + 2x 1 + … + 2 N – 1 ⋅ x N – 1 (1-5)
k=0
Using the offset binary code we cannot represent negative numbers. The range of the
code is
0 ≤ X ≤ 2N – 1 (1-6)
In the thesis we focus on using offset binary codes. In current-steering DACs this is a
convenient method, Chapter 4 in Part B: Nyquist-Rate Digital-to-Analog Converters for
VDSL Applications .
The signed-magnitude code uses the MSB as sign bit. x N – 1 = 0 indicates a positive
number and x N – 1 = 1 a negative. The other bits indicate the magnitude as for an offset
binary code. A number is given by
N–2
X = ( –1 ) xN – 1
⋅ ∑ 2k ⋅ xk (1-7)
k=0
One effect is that there are two representations for the zero value; 0…00 and 10…00 .
The signed-magnitude code also may have operand sign-dependent operations which
introduces extra control logic and computation time [42]. The range of the code is
X ≤ (2N – 1 – 1) (1-8)
The signed-magnitude allow though some special structures for binary weighted current-
steering DACs [56].
In a one’s complement code we define the number as
N–2
X = –xN – 1 ⋅ (2N – 1 – 1) + ∑ 2k ⋅ xk (1-9)
k=0
Also with this representation there is a situation with two zeros; 0…00 and 1…11 . The
10
.
X ≤ (2N – 1 – 1) (1-10)
The two’s complement code also utilizes a sign bit in the MSB position. The digital
number is given by
N–2
X = – 2N – 1 ⋅ xN – 1 + ∑ 2k ⋅ xk = x 0 + 2x 1 + … – 2 N – 1 ⋅ x N – 1 (1-11)
k=0
–2 N – 1 ≤ X ≤ ( 2 N – 1 – 1 ) (1-12)
The two’s complement code has good properties for arithmetic operations and therefore
used in several applications.
In the thermometer code, the number of ones in the code determines the represented
number. The number is given by
2N – 1
X = ∑ xk (1-13)
k=1
The thermometer code is suitable for segmentation of sources, when a number of equally
large sources is desired.
11
Part A: Overview: Digital-to-Analog Converter Design
Broadband Splitter
Network Local
e.g., Radio, Network
Fibre, etc. “existing”
copper wires
Telephone
(Voice)
Figure 1-7: Applications of the xDSL technique. The information from a broadband network is
transferred to the local network via copper wires of the existing telephone net.
Since the transmission speed for ADSL can be around 6Mb/s, it is a suitable standard for
internet access, services as video-on-demand and similar. For VDSL the transmission
rates are even higher and therefore it may be used for broadband network.
ADSL
POTS
HDSL
VDSL
f [Hz]
4.3k 1.1M 50M
Figure 1-8: Frequency domain characteristics of different xDSL techniques.
The frequency ranges for different standards are shown in Fig. 1-8 and summarized in
Table 1-2.
12
.
The xDSL standards use modulation techniques as for example discrete multi-tone
(DMT), carrierless amplitude phase (CAP), and discrete wavelet multi-tone (DWTM).
The DMT technique is basically a version of orthogonal frequency division multiplexing
(OFDM). We briefly discuss the standards; ADSL and VDSL:
• Asymmetric digital subscriber’s line (ADSL)
The ADSL uses an asymmetric transmission scheme, since for most applications from
the single user’s point of view, the downstream transmission speed is higher than the
upstream, which is utilized in ADSL. With the DMT modulation the data is associated
with a number of carriers. The carriers are individually orthogonal and are at the positions
fk = k ⋅ f0 (1-15)
where f 0 = 4.3125 kHz and k is an integer. The number of carriers used is determined
by the quality of the line and the mode of operation. If the line is perfect 249 tones are
used in the frequency range from 25.875kHz to 1.104MHz. The output spectrum of a
DMT signal with 240 tones is shown in Fig. 1-9. Shown in the figure is also the frequency
domain specification according to the ANSI T1E1.4 working group [83] as summarized
in Table 1-3.
ADSL Specifications
−40
−60
−80
−100
−120
−140
−160
−180
4 5 6 7
10 10 10 10
Figure 1-9: Output amplitude spectrum of a 240-tone DMT signal shown with frequency domain
specifications for the ADSL standard [83].
The number of bits associated with a certain carrier is given by the signal-to-noise ratio
within a 4.3125kHz bandwidth centered around the specific carrier. According to the
specifications [82, 83], the SNDR has to be given (in dB) by
SNDR ≥ 3N + 20 and SNDR ≥ 38 (1-16)
where N is the number of bits that can be associated with the carrier.
Since the ADSL specifications still are hard to meet, the ADSL-Lite standard (Table 1-
2) with reduced requirements has evolved. The reader can find more information about
the overall ADSL techniques in [82-90].
13
Part A: Overview: Digital-to-Analog Converter Design
a) D b) S
B G B G
S D
Figure 1-10: Circuit symbols for the a) NMOS and b) PMOS transistor.
For the PMOS transistor we use heavily p-type doped source and drain. The substrate is
of n-type. For the NMOS transistor the source and drain are heavily n-type doped, and the
substrate is p-type doped.
14
.
In weak inversion (subthreshold region) the transistor is cut-off and only a very small
current (leakage current) is flowing through the transistor. In strong inversion the gate-
source voltage, v GS , is higher than the threshold voltage, v T , and a considerably larger
current, i D , may flow from source to drain. When the drain-source voltage, v DS , is lower
than the effective voltage, v eff = v GS – v T , the current is linearly dependent on the gate-
source voltage. When the drain-source voltage, v DS , is higher than the effective voltage,
v eff , the channel will become pinched-off, i.e., saturated, and the current is no longer lin-
early dependent on the gate-source voltage. The drain current, i D , for the NMOS transis-
tor in the different operating regions is approximately given by [37]:
• Cut-off region, ( v GS < v T ) :
iD ≈ 0 (1-17)
µ 0 C ox W
i D ≈ --------------- ⋅ ----- ⋅ ( 2 ( v GS – v T ) – v DS ) ⋅ v DS ⋅ ( 1 + λv DS ) (1-18)
2 L
µ 0 C ox W
i D ≈ --------------- ⋅ ----- ⋅ ( v GS – v T ) 2 ⋅ ( 1 + λv DS ) (1-19)
2 L
where λ is the channel length modulation factor, µ 0 is the charge mobility, C ox is the
oxide capacitance per gate area, W is the channel width and L is the channel length. The
threshold voltage, v T , is given by
v T = V T 0 + γ ⋅ ( 2 φ F + v SB – 2 φ F ) (1-20)
where V T 0 is a constant given by the material, γ is the body-factor, and 2 φ F is the sur-
face potential at strong inversion. The notations
15
Part A: Overview: Digital-to-Analog Converter Design
∂i D
g ds = ----------
- (1-24)
∂v DS
Q
The influence of the source-bulk voltage also gives rise to a gain factor of
∂i D ∂v T
g mbs = ----------
- = g m ⋅ ----------- (1-26)
∂v BS ∂v SB
Q Q
16
.
bandwidth of the circuit is affected. In Table 1-4 the capacitor sizes in different operation
regions are summarized.
Cgs Cgd
S D
Csb Cgb Cdb
B
Figure 1-12: Capacitors associated with the terminals of the transistor, Cgs, Cgd, Csb, Cgb, and Cdb.
Cut-off C ox WL 0 0
Linear 0 C ox WL ⁄ 2 C ox WL ⁄ 2
Saturation 0 2C ox WL ⁄ 3 0
Table 1-4: Variations of capacitors associated with the transistor in different working conditions.
The charges at C gs and C gd are changing when the switching voltage at the gate is
changing. This forces extra charge to redistribute to the drain and source. When switching
analog signals using digitally controlled switches, e.g., sample-and-hold circuits, the
capacitors may introduce glitches and current spikes in the analog signal. This clock
feedthrough (CFT) gives rise to noise and distortion in the analog signal [71].
i2
S ( f ) = ------ (1-28)
∆f
where i 2 is the average normalized mean-square noise current and ∆f is the frequency
variable. The shot noise, thermal noise, and 1 ⁄ f noise (flicker noise) are briefly dis-
cussed:
• The shot noise PSD, arising when a dc current, I , is applied to a pn-junction, is
given by
S ( f ) = 2qI
where q is the fundamental electron charge.
• The thermal noise PSD through a resistance, R , is determined by
4kT
S ( f ) = ----------
R
17
Part A: Overview: Digital-to-Analog Converter Design
where k is the Boltzmann’s constant and T is the absolute temperature. For the
CMOS channel we have an extra correction factor to the expression. The PSD of
the noise current through the drain of the transistor is given by
8
S ( f ) = --- kT g m
3
where g m is the transconductance of the transistor.
• The flicker noise (or 1 ⁄ f noise) PSD is determined by
Ia
S ( f ) = K f ⋅ -----b
f
where a , b , and K f are material-dependent variables. For a and b , typical values
are a ≈ 1 and 0.8 ≤ b ≤ 1.3 .
For high-speed CMOS applications it is the thermal noise that is dominating. For dis-
crete-time circuits we also have to consider the concept of sampled noise or switched
noise.
18
2. The Ideal DAC
X ( e jωT ) = X ( e j ( ωT + 2π ) ) (2-2)
The amplitude spectrum is shown in Fig. 2-1. For a discrete-amplitude signal, as dis-
cussed above, there would be a noise floor, or even distortion terms, through out the fre-
quency range as well.
|X(ejωT)|
ejωT
−2π −π π 2π
Figure 2-1: Discrete-time amplitude spectrum.
The generation of the continuous-time signal, hence the extraction of the wanted infor-
mation, from the discrete-time spectrum is performed with an ideal lowpass filter as
shown in Fig. 2-2. The frequency spectrum of the continuous-time signal, X ( ω ) , is
19
Part A: Overview: Digital-to-Analog Converter Design
|X(ejωT)|
ejωT
−2π −π π 2π
Figure 2-2: Ideal lowpass filtering with pulse amplitude modulation.
where the discrete-time values are weighted with the sinc functions, i.e., pulse amplitude
modulation (PAM). This is however impossible to accomplish in a real implementation
since the sinc function is non-causal. In the real case, the common way to reconstruct the
signal is to use sample-and-hold circuits on the output, hence the output is described by
pulse waves. Let x̂ ( t ) be the reconstructed output signal as
x̂ ( t ) = ∑ x ( nT ) ⋅ [ u ( t – nT ) – u ( t – ( n + 1 )T ) ] (2-5)
∀n
where x ( nT ) are the discrete-time values and u ( t ) is the step function. The frequency
spectrum of the reconstructed signal is
1 – e – jωT
X̂ ( ω ) = F { x̂ ( t ) } = ---------------------- ⋅ ∑ x ( nT ) ⋅ e – jωnT =
jω ∀n
T
– jω ---
= e 2 ⋅ sinc ( ωT ⁄ 2 ) ⋅ X ( e jωT ) = C ( ω ) ⋅ X ( ω ) (2-6)
The amplitude spectrum of the sinc weighting function, C ( ω ) , is shown in Fig. 2-3. As
is seen from the figure there is an attenuation of the original signal spectrum within the
Nyquist frequency band, f N = f s ⁄ 2 . Images also occurs since C ( ω ) is non-zero at fre-
quencies above the Nyquist frequency and the discrete-time signal is repeating, (2-2). The
images are removed or attenuated with an LP filter with cut-off frequency at the Nyquist
frequency. To compensate the attenuation within the signal band, an anti-sinc filter at the
output may be used as well. At the Nyquist frequency, f N , there is an attenuation of
C ( 2π ⋅ f N ) = C ( π f s ) = 2 ⁄ π ≈ 0.637 (2-7)
A simple approach to compensate the distortion is to use a filter, H as , that fulfils the
approximate equation
H as ( ω ) ⋅ C ( ω ) ≈ 1 (2-8)
fs
in the frequency range ω ≤ 2π ⋅ ----- .
2
The filter can be implemented in the digital domain at the input of the digital-to-analog
conversion as well as using an analog filter at the output [7, 17, 30, 47].
20
2. The Ideal DAC
|C(ω)|
1
0.64
ω
πfs 2πfs 3πfs 4πfs
Figure 2-3: Sinc weighting characteristics due to sample-and-hold elements.
X ∈ { X 0, X 1, …, X ( 2 N – 2 ), X ( 2 N – 1 ) } (2-9)
Let the function describing the transfer function of the converter be given by A ( X )
where A ( X ) is the amplitude level generated for the code X . For a specific code, X i , we
denote the amplitude as
A ( X i ) = Ai (2-11)
where A ref is the reference level and i ≤ 0, …, 2 N – 1 . Further, the reference level is the
amplitude level for the LSB and the notation
21
Part A: Overview: Digital-to-Analog Converter Design
is also used. As described in equation (1-1) the output amplitude for a binary weighted
converter is given by
where the result from (2-12) is applied. The maximum output amplitude is given by
Assuming that the input signal is statistically equal distributed over the whole amplitude
range, the uniform quantization intuitively provides a minimum error, hence a minimum
quantization noise.
A(Xi) A(Xi)
a) b)
Ak+1 Ak+1
Ak X Ak X
When the signal is not equally distributed over the amplitude range it may be more useful
to use non-uniform quantization, Fig. 2-4b). With this method some amplitude levels are
quantized with a larger error, but statistically a minimum quantization noise is achieved
since the probability for these amplitude levels is lower. The output level using non-uni-
form quantization can be written as
A ( X ) = x0 ⋅ A0 + x1 ⋅ A1 + x2 ⋅ A2 + … + x N – 1 ⋅ A N – 1 (2-16)
22
2. The Ideal DAC
where A k = k ⋅ A ref is the amplitude level for the code X k , and u ( t ) is the step function.
The wanted output signal, A want ( t ) , should be a ramp function, Fig. 2-5, as
t t t
A want ( t ) = --- ⋅ A max = --- ⋅ ( A ref ⋅ 2 N ) = ------- ⋅ A ref (2-18)
T T ∆T
‘
Aout
8Aref Awant
7Aref Atrue
6Aref
5Aref
4Aref
3Aref
2Aref
Aref
X
000 001 010 011 100 101 110 111
Figure 2-5: Transfer characteristics for an offset binary conversion and uniform quantization.
The quantization error as function of the time, shown in Fig. 2-6, is written as
We can now find the error r.m.s. value, A err, rms , by using
1 T
T ∫0 want
2
A err , rms = --- [A ( t ) – A true ( t ) ] 2 dt (2-20)
It can easily be seen that the total r.m.s. error (over the time T ) must be the same as the
r.m.s. error during half a single time period of ∆T ⁄ 2 , therefore (2-20) can be written as
2 ∆T ⁄ 2 2
A ref t 2 A ref
2
A err , rms = --------------
∆T ⁄ 2 ∫ ------- dt = ---------
∆T 12
- (2-21)
0
23
Part A: Overview: Digital-to-Analog Converter Design
A ref
A err, rms = ---------
- (2-22)
12
a) Aerr(t) T
t
A ref
– ---------
-
2
Aerr(t)
b)
Awant(t) Atrue(t)
Figure 2-6: Quantization error as a) function of time and b) modeled as an error signal. Awant(t) is the
wanted signal, Atrue(t) is the true output, and Aerr(t) is the quantization noise.
1 1
A sig, rms = ------- ⋅ A sig, ampl = ------- ⋅ 2 N – 1 ⋅ A ref (2-24)
2 2
We define the peak signal-to-noise ratio (SNR) as
24
3. DAC Performance Measures
When the code is switching, clock feedthrough (CFT), glitches, overshoot, etc., may
arise. CFT arises due to the capacitive coupling between the digital switching signal and
the analog output signal [71]. The glitch depends on how the internal switches are indi-
vidually skewed. It may be an erroneous code connected to the output for a short period
of time. This is further described in Sec. 3.2.
25
Part A: Overview: Digital-to-Analog Converter Design
By using the known deviations, d i , we can characterize the performance and behavior
of the converter.
di
Ai Ts t
Figure 3-1: True output signal compared with wanted output signal (dashed). The deviation from the
start-value, di, and the end-value, di+1, determine the static value of the specific code.
Settling time is given by Ts.
For a large number of bits, it may be inconvenient to find all static values by pure mea-
surement, since the number of amplitude levels grows exponentially. Therefore statistical
methods can be used to find the deviation values and use them to derive static measures
[16].
A want ( X i ) = A i (3-1)
The true value generated by the converter for the same code is given by
A true ( X i ) = A˜ i (3-2)
The deviation is normalized with respect to the LSB value, A LSB from (2-13), as
di A˜ i – A i
d i = -----------
- = ---------------- (3-4)
A LSB A LSB
and the deviation gets the unit “LSB”. The DNL, D i , is the difference between the two
deviations at a certain code transition, X i – 1 → X i , as
A˜ i – A i à i – 1 – A i – 1 à i – à i – 1
D i = d i – d i – 1 = ---------------- – ------------------------------ = ----------------------
-–1 (3-5)
A LSB A LSB A LSB
26
3. DAC Performance Measures
Note that in this case, A i and A i – 1 are the amplitude levels of two consecutive codes
and the difference between them is ideally one LSB.
Aout
7Aref
6Aref
5Aref
4Aref d4
A4
3Aref
2Aref
Aref
X
000 001 010 011 100 101 110 111
Figure 3-2: Possible errors in amplitude levels for a 3-bit DAC.
If all deviations are equally large, the DNL is D i = 0 for all code transitions, and there
will only be an offset error at the output. A basic observation is that if the DNL is increas-
ing, the deviations from the true code is also increasing and the transfer function is non-
linear.
In most configurations the converter is constructed with differential outputs where offset
errors (between two channels) are cancelled, and it may in that case be useful to define
the offset value, A offset , as the average value of all deviations. The offset value is found
by minimizing the expression d i – A offset for all i , with the least square method as
2N – 1
∂ 1
∂ A offset
-----N-
2 ∑ ( d i – A offset ) 2 = 0 (3-7)
k=0
which gives
2N – 1
1
A offset = -----N- ⋅
2 ∑ di (3-8)
k=0
27
Part A: Overview: Digital-to-Analog Converter Design
and the DNL values are not affected by this modification, see (3-5).
Aout Aout
Atrue(X) Atrue(X)
Awant(X) Awant(X)
X X
Linear gain error does not introduce extra distortion as long as the output signal is not
clipping or similar effects arise, compare with (2-14). The true output with a different
slope and with offset error can be written as
A true ( X ) = a ⋅ X + A offset (3-10)
The non-linear errors may be reduced by using pre-distortion [54], see Sec. 5.1.1 in
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications.
i
Ii = ∑ Dk (3-12)
k=1
28
3. DAC Performance Measures
à i – A˜0 i ⋅ A LSB
i
à i – à i – 1 di – d0
Ii = ∑ ----------------------- – 1 = ----------------- – ------------------ = ---------------
A LSB A LSB A LSB A LSB
- (3-13)
k=1
As was discussed in the previous sections, offset errors and linear gain errors may be
neglected since they do not necessarily degrade the performance. Therefore, compensa-
tion for offset and linear gain error gives a more understandable value of the INL and
DNL. The wanted straight line can be defined in two ways, either using the start and end
values as reference or using a best-fit straight line, which minimizes the error. The first
guarantees a true zero code definition and the latter is more useful when using differential
structures.
The deviation values are compensated as (3-9) but now the gain error, a , is included as
well
The gain and offset errors, a and A offset , are found by minimizing and using the least
square method as
2N – 1 2N – 1
∂ 1 ∂ 1
∑ ∑ [ A˜ i – ( A offset + a ⋅ k ) ] 2 = 0
2
------ dˆi = ------ (3-15)
∂a 2N ∂a 2N
k=0 k=0
and
2N – 1
∂ 1
∂ A offset
-----N-
2 ∑ [ A˜ i – ( A offset + a ⋅ k ) ] 2 = 0 (3-16)
k=0
If we assume a large number of bits, the following approximate values are found
2N – 1 2N – 1
4 6
A offset ≈ -----N-
2 ∑ A˜k – --------
2 2N
- ∑ k ⋅ A˜k (3-17)
k=0 k=0
2N – 1 2N – 1
12 6
a ≈ --------
2 3N
- ∑ k ⋅ A˜k – --------
2 2N
- ∑ A˜k (3-18)
k=0 k=0
The line A offset + a ⋅ X expresses the best-fit straight line with respect to the true output
values, A˜ i . The compensated values, dˆi from (3-14), are used to find the DNL and INL
values as was given in equations (3-5) and (3-12).
3.1.5 Monotonicity
If the output amplitude level of the converter is increasing with increasing input and if
there is a one-to-one mapping from the input to the output, the converter is monotonic.
Otherwise there is a situation when a specific output value may be represented by two dif-
ferent input codes, see Fig. 3-4.
29
Part A: Overview: Digital-to-Analog Converter Design
1 LSB
di+1
di
Figure 3-4: Transfer function illustrating when monotonicity is not fulfilled (solid line).
Monotonicity implies that all DNL values have to be less than one LSB [56] as
D i ≤ 1 LSB for all i (3-19)
This also implies that the deviation from the best-fit straight line must never be more
than a half LSB [56], hence the converter is monotonic if the INL values fulfil the ine-
quality
1
I i ≤ --- LSB for all i (3-20)
2
Monotonicity can be improved by segmenting the bits, i.e., using a thermometer code.
This is further discussed in Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL
Applications.
3.2.1 Glitch
Glitches occur when the time of switching between different bits in the converter is
unmatched. For a short period of time a false code could be represented at the output. For
example if the code transition is
0111…111 → 1000…000
If the MSB is switching faster than the LSBs, the code 11…111 may be present for a
short time. This code represents the maximum value and hence the glitch would be large.
It is important to understand how the glitches may affect the performance. We use glitch
energy as a measure of their impact.
We model the glitch as a pulse, Fig. 3-5, with a certain amplitude height, A g , and with
a time duration, T g . The normalized average power, P g , of the glitch distributed over the
30
3. DAC Performance Measures
shortest possible code duration, i.e., the update time interval, T s = 1 ⁄ f s , is described by
Tg
P g = A g2 ⋅ ------ (3-21)
Ts
Ag
Aj
Tg
Ak
t
Figure 3-5: Glitch modeled as a pulse with height Ag and duration Tg.
Assume that the maximum peak glitch amplitude, i.e., the amplitude of the MSB, is
which gives the maximum glitch power over one clock cycle
Tg
P g, max = 2 2N – 2 ⋅ A LSB
2 ⋅ ------ (3-23)
Ts
This should be compared with the power of the quantization noise, P Q , (2-21), during
the same period of time
2
A LSB
P Q = -----------
- (3-24)
12
The power of the glitch should be smaller than the quantization in order to not decrease
the SNR too much, hence
T g A LSB 2
P g, max = 2 2N – 2 ⋅ A LSB
2 ⋅ ------ < ------------ = P Q (3-25)
Ts 12
31
Part A: Overview: Digital-to-Analog Converter Design
The CFT error may be modeled in a similar way as was done for glitches. Naturally, the
CFT is reduced when reducing the capacitive coupling. The transistor sizes may be
reduced to decrease the size of the parasitic capacitances. With smaller transistors, the on-
resistance increases and may degrade the result in other ways, i.e., voltage drop over the
switch, etc.
t
A out ( t 0 + t ) = A out ( t 0 ) + A step ⋅ 1 – e RC
– --------
(3-27)
where A step is the wanted step and τ = RC is the time constant determined by the resis-
tance and capacitance associated with the output of the converter. R and C are both de-
pendent on the input code (and the output amplitude as well). Assume that the true output
value should be found within one clock period, T s . The error at the output, d ( T s ) , is giv-
en by
T
A out ( T s ) – A out ( t 0 ) – -------s-
d ( T s ) = ----------------------------------------------- = 1 – e RC
A step
For example, for an error less than 1%, d ( T s ) < 0.01 , the clock period must be
T s > τ ⁄ 100
This indicates the maximum speed of the circuit and it also gives a guide how to design
the circuit according to the specifications. It is further discussed in Sec. 3.5 in Part B:
Nyquist-Rate Digital-to-Analog Converters for VDSL Applications.
32
3. DAC Performance Measures
where A rms is the signal r.m.s. value and B rms is the noise r.m.s. value. With a full-scale
sinusoidal signal and quantization noise only, the SNR was derived in section 2.2 to be
SNR ≈ 6.02 ⋅ N + 1.76 (3-29)
The harmonic distortion with respect to the k -th harmonic ( HD k ) is the power ratio
between the k -th harmonic and the fundamental as
A k2 ⁄ 2 A k2
HD k = ------------
- = -----
- (3-30)
A 02 ⁄ 2 A 02
where A k is the amplitude of the k -th harmonic and A 0 is the fundamental’s amplitude.
A k2 1
THD = ∑ HD k = ∑ A 2- = -----
-----
A 2 ∑ k
-⋅ A2 (3-31)
k≥1 k≥1 0 0 k≥1
where A k is the amplitude of the k -th harmonic and A 0 is the fundamental’s amplitude.
A 02 ⁄ 2 SNR
SNDR = ------------------------------------- = ------------------------------------- (3-32)
1 1 + THD ⋅ SNR
--- ∑ A k2 + B rms 2
2k ≥ 1
33
Part A: Overview: Digital-to-Analog Converter Design
where A 0 is the amplitude of the fundamental, A k is the amplitude of the k -th harmonic,
and B rms is the noise r.m.s. value.
A 02 ⁄ 2
SFDR = ------------
- (3-33)
P max
where A 0 is the amplitude of the fundamental and P max is the power of the second largest
component. The SFDR mostly determines the quality of the converter. Assume that the
maximum peak is one of the harmonics, P max = A max 2 ⁄ 2 . Then (3-31) is rewritten as
2
A max A k2
THD = -----------
A 02
- ⋅ 1 + ∑ 2 - ----------- (3-34)
k ≠ maxpos A max
If the sum in expression (3-34) is small enough, the THD is determined by the SFDR.
Using (3-32) we also see that if the distortion is larger than the SNR, also SNDR will be
determined by the SFDR.
where k and m are integer numbers, and further k ≠ 0 , m ≠ 0 , and f 1 ≠ f 2 . The inter-
modulation distortion is found by
IMD =
∑ A k2, m
------------------- (3-37)
A 02
where A 0 is the amplitude of the fundamental and A k, m is the amplitude of the tones in
34
3. DAC Performance Measures
A2 ⁄ 2
MTPR = ------------------------------------------------------------------------------------------------------------- (3-38)
Power of tones at left-out frequency positions
This is also depicted in Fig. 3-6, where 25 tones have been applied. At two frequencies
tones have been excluded. The non-linearity of the converter introduces harmonics at
these positions, and MTPR can be determined by the power of these harmonics.
80
60
40
MTPR
20
Signal power.
−20
−40
−60
−80
−100
−120
0 0.1 0.2 0.3 0.4 0.5
Normalized frequency.
Figure 3-6: Illustration of the impact of MTPR. Two tones are left out and the non-linearity can be
found by observing the power of the tones appearing at the left-out positions.
35
Part A: Overview: Digital-to-Analog Converter Design
2
where A max is the normalized peak power and A rms is the signal r.m.s. value. For a si-
nusoidal the PAR = 2 since A max = A 2 and A rms
2 = A 2 ⁄ 2 where A is the amplitude
of the sinusoidal.
Resolution
(bits)
20 Oversampling DACs
Current-Mode
Nyquist-Rate DACs
12
Voltage-Mode
Basically, it can be seen that Nyquist-rate converters are the most suitable converter
structures for high-speed telecommunication applications. For audio applications and
lower-speed telecommunication applications the oversampling converter is a suitable
structure. The current-mode Nyquist-rate versions are suitable for high-speed applications
[2, 3, 9, 13, 30-32, 39, 44, 49, 53, 61, 64-66, 68, 70] and voltage-mode oversampled ver-
sions for lower-speed applications [40, 46, 51, 59, 62, 79].
36
Nyquist-Rate
B
Part B:
Digital-to-Analog
Converters for
VDSL Applications
1 Introduction
The Nyquist-rate digital-to-analog converter is defined to have its signal frequency range
from zero (dc) up to the Nyquist frequency, f N = f s ⁄ 2 , with f s as the update frequency
of the DAC [37, 56]. In Fig. 1-1 the amplitude spectrum is shown where the sinc weight-
ing from sample-and-hold elements has not been considered. Images of the signal are also
found in the spectrum.
|X(f)| Image
Quantization
Noise
fs /2 fs f
Figure 1-1: Amplitude spectrum within the Nyquist frequency range. The signal’s images are situated
above the Nyquist frequency. The quantization noise is equal over all frequencies.
The quantization noise within the Nyquist band determines the signal-to-noise ratio as
discussed in the previous part of the thesis. With a normalized full-scale input sinusoidal
37
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
(single-side peak at 0dB) the average quantization noise power, Q avg, pwr , is found at a
level of
From (1-1) we have the intuitive conclusion, that the higher bandwidth, f N , the lower
noise floor. The output from the converter when using sample-and-hold circuits also has
images and attenuation of the signal spectrum due to the sinc weighting in the frequency
domain as sketched in Fig. 1-2. Lowpass filters and anti-sinc filters can be used to remove
images and reduce the distortion.
|X(ω)|
1 1st Nyquist
band Images
fs/2 fs 2fs f
Figure 1-2: The output amplitude spectrum with images and the effect of sinc weighting.
In general the Nyquist-rate converter is required for wideband applications, e.g., wide-
band radio, VDSL, etc., where oversampling techniques are impossible due to the high
clocking frequencies needed.
A brief description of different DAC structures is given in Chapter 2. The R-2R ladder,
charge-redistribution, and current-steering structures are discussed. This thesis’ part con-
centrates on the CMOS binary weighted current-steering DAC which provides a simple
and straight-forward structure suitable for high-speed telecommunication applications as
mentioned in Sec. 3.4 in Part A: Overview: Digital-to-Analog Converter Design.
In Chapter 3 we model the performance of a current-steering DAC. Important aspects
concerning output impedance, matching, and noise are discussed. It is very important to
have a knowledge about the overall behavior of the circuit and how irregularities and dif-
ferent circuit variations will affect the result. Therefore, the modeling results are used to
iterate the design of the converter using behavioral-level descriptions.
The design of a DAC chipset is presented in Chapter 4. Behavioral-level descriptions and
simulations concerning the design is given, as well as results from circuit simulations and
measurements.
From the design of the chipset as well as the modeling it can be concluded that several
different improvements may be done. Some of these are discussed in Chapter 5.
38
2. Nyquist-Rate DAC Structures
φ1
16C
φ2
8C 4C 2C C v0(t)
φ1
x3 x2 x1 x0
Vref φ1
C0
φ2
φ2
This switched capacitor version uses two non-overlapping clock phases, φ 1 and φ 2 . The
operation is determined by the discrete-time events controlled by the clocks. During the
39
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
initializing phase; time point nT – T ⁄ 2 and clock phase φ 2 , all inputs are shorted and no
charge is found on the binary weighted capacitors. T = 1 ⁄ f s is the update time. The
16C capacitor is charged with q 16 ( nT – T ⁄ 2 ) = 16C ⋅ v 0 ( nT – T ⁄ 2 ) and C 0 is
charged with q 0 ( nT – T ⁄ 2 ) = C 0 ⋅ v 0 ( nT – T ⁄ 2 ) .
During clock phase φ 1 and at time point nT , the binary weighted capacitors are charged
with the reference voltage, V ref . The switches controlled by the digital input
X = ( x 3, x 2, x 1, x 0 ) , determine what capacitors to charge, i.e., the total charge on those
capacitors is given by q tot ( nT ) = X ( nT ) ⋅ C ⋅ V ref
Since the 16C capacitor is discharged and C 0 is connected to the input of the opera-
tional amplifier at this point, they will provide the binary weighted capacitors with charge.
Charge conservation at the virtual ground state that
X ( nT ) ⋅ C ⋅ V ref + C 0 ⋅ v 0 ( nT ) = 16C ⋅ v 0 ( nT – T ⁄ 2 ) + C 0 ⋅ v 0 ( nT – T ⁄ 2 )
After one clock period, during clock phase φ 2 and at time point nT + T ⁄ 2 , the binary
weighted capacitances are discharged and their charge is redistributed to the 16C capac-
itor. This implies that
q 16 ( nT + T ⁄ 2 ) = 16C ⋅ v 0 ( nT + T ⁄ 2 ) = X ( nT ) ⋅ C ⋅ V ref
40
2. Nyquist-Rate DAC Structures
allel, hence the MSB current source therefore has the dc value I MSB = 2 N – 1 ⋅ I LSB .
Vref
IMSB ILSB
xN-1 xN-2 x0
Iout
Figure 2-2: An N-bit current-steering DAC.
The output current, I out , of the DAC shown in Fig. 2-2 is given by
The current-steering DAC has the advantages of being quite small for resolutions below
10 bits and it may be very fast. The major disadvantage is its sensitivity to device mis-
match and glitching. To increase monotonicity and reduce the influence of glitches, a ther-
mometer code can be used. This is further discussed in Sec. 5.2.1. The current-steering is
suitable for high-speed wideband applications when special care is taken to improve the
matching of the converter.
Results from modeling of the current-steering DAC is given in Chapter 3 where the fun-
damental impacts of matching error, output impedance, noise, etc., are discussed.
If the input is zero, X = 0 and hence all x i = 0 , it is seen that the output resistance of
the ladder is 2R . When terminating the output with 2R to the V ref voltage, the wanted
transfer function is achieved and by determining the contribution from each source, the
operation of the converter is easily understood.
Due to the fact that the current sources are equally large, the matching can be improved.
The resistors are however often non-linear and contains signal-dependent capacitances
yielding distortion. The ladder also introduces a time-delay between the MSB and LSB
which generates glitches and distortion. There is the same amount of current through all
41
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
switches which makes the design of the switches simpler, however, the internal voltage
nodes are swinging and therefore the current sources will have varying terminal voltages,
hence a poorer matching, further introducing distortion.
Vref
R 2R 2R
R R R Iout
x0 x1 xN-1
I0 I0 I0
Vss
Figure 2-3: An N-bit R-2R ladder DAC.
One of the major advantages is that we only have a few number of different component
types to implement, i.e., two resistor sizes, R and 2R , one current source size, I 0 , and
one type of current switch. This allows a more regular layout and since the current sources
all have the same size, trimming or calibration can be applied [8, .
2.4 Comparison
The properties of the different DAC structures described in Sec. 2.1 through Sec. 2.3 are
approximately given in Table 2-1. The values are found from reported results in literature
and data sheets from vendors [2, 3, 9, 13, 30-32, 39, 44, 49, 53, 61, 64-66, 68, 70].
One way to improve the overall performance is to combine different DAC structures into
a hybrid DAC. Naturally a large number of hybrid DACs can be found. For example the
LSBs can be converted using a current-steering structure and the MSBs are converted
with an R-2R structure, since the impact of matching errors are larger on the MSBs. Fur-
ther improvements are discussed in Chapter 5.
42
3. DAC Performance Modeling
W/L
Rout
Iout
Figure 3-1: Illustration of conflicts when choosing design parameters in a current-steering DAC.
43
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
Linearity is affected by the output impedance R out , the noise is affected by the output
current, I out , and matching is practically determined by the transistor size W ⁄ L or gate
area WL . There is however a relation between the three properties and a trade-off has to
be done to find a solution that is meeting the specifications. When increasing the output
current with the ambition to improve the SNR, the output impedance is decreasing and
thereby the non-linearity increase, etc.
Output Level
Aout
Wanted transfer function
7Aref
Actual transfer function
6Aref
5Aref
4Aref
Errors or deviations
3Aref
2Aref
Aref
Code X
000 001 010 011 100 101 110 111
Figure 3-2: Wanted transfer characteristics (dashed line) and with example of errors in different
conversion levels caused by the actual transfer characteristics (solid line).
If one assumes that the level errors of the transfer function caused by matching errors are
frequency-independent (which is not the case) the total error at the output is signal-depen-
dent as well. This implies the non-linearity of the transfer function. Mainly, the SFDR and
the SNR are used to characterize the converter. In principle, the SFDR determines the
THD, SNDR, and MTPR if the converter is not too non-ideal, i.e., not too much noise and
not too non-linear, compare with the formulae given in Part A: Overview: Digital-to-Ana-
log Converter Design.
The SFDR can be deduced from DNL and INL values if the input signal is known, and
especially how the input signal is sweeping over all transfer function levels. If the input
signal is not full-scale, all DNL values will not affect the linearity of the output signal, etc.
Basically, current sources used in a current-steering DAC is constructed by using a num-
ber of unit current sources (binary weighted) in parallel. This property is used throughout
the following sections.
44
3. DAC Performance Modeling
VDD
IS(X)
RS(X)
RL
IL
Figure 3-3: Current source, IS(X), with finite output resistance, RS(X). The termination, RL, is signal-
independent.
We only show the impact of output resistance, since the influence of capacitive elements
strongly affects the complexity of the modeling (Sec. 3.5). The finite output resistance
reduces the current delivered to the load. The current delivered to the load from a non-
ideal current source as well as the supply voltage V DD is given by
RS 1
I L = ------------------- ⋅ I S + ------------------- ⋅ V DD (3-1)
RS + RL RS + RL
where I S is the nominal current value in the current source given by (2-1), R S is the out-
put resistance of the current source, and R L is the load (or termination) resistance.
From (3-1) it is seen, that if the output resistance were constant, there would only be a
loss of gain and a dc current in the output signal, which would not degrade the perfor-
mance. However, as was discussed earlier, the total output resistance is dependent on the
number of sources connected to the output. This dependency give rise to distortion. Using
a notation describing the code dependency, (3-1) can be rewritten as
I S ( X ) + G S ( X ) ⋅ V DD
I L ( X ) = ---------------------------------------------------
- (3-2)
1 + GS ( X ) ⋅ RL
where X indicates the DAC’s digital input given by (2-2) and G S ( X ) = 1 ⁄ R S ( X ) is the
code-dependent output conductance of the sources. It will be shown that this affects the
second harmonic strongly for reasonable high ratios of R S ⁄ R L , for lower ratios also high-
er order harmonics will be large in the output signal spectrum.
The termination resistance, R L , is assumed to be fixed and signal-independent. The out-
put resistance, R S , and the output conductance, G S , are determined by studying the
DAC’s structure shown in Fig. 2-2. The output conductance, G S ( X ) , is related to the
45
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
number of parallel current sources that are currently connected to the output node. The
output conductance associated with the LSB current source is assumed to be G LSB and
therefore the output conductance of the i -th LSB must be given by G i = 2 i – 1 ⋅ G LSB .
The total output conductance, G S ( X ) , of the converter is given by
where X is the input code (offset binary code) given by (2-2), and N is the number of bits.
Combining (2-1), (3-2) and (3-3), we have
I LSB + G LSB ⋅ V DD 1
I L ( X ) = --------------------------------------------- ⋅ 1 – ----------------------------------------- (3-4)
G LSB ⋅ R L 1 + X ⋅ G LSB ⋅ R L
Assume that within the Nyquist band, the input signal, X , is a full-scale sinusoidal with
an offset and can be written as
X = X ( k ) = A ⋅ [ 1 + sin ( Ωk ) ] + e (3-5)
where e is the deviation from an ideal sinusoidal due to truncation, i.e., quantization
noise, Ω is the normalized angular frequency, k is a sequence index, and the amplitude,
A , is given by
2N – 1
A = --------------- (3-6)
2
since the input is an offset binary code. Combining (3-5) with (3-4) gives the expression
of the current through the load resistance due to the current division.
I LSB + G LSB ⋅ V DD 1
I L ( k ) = --------------------------------------------- ⋅ 1 – -------------------------------------------------------------------------------- (3-7)
G LSB ⋅ R L 1 + G LSB ⋅ R L ⋅ A ⋅ [ 1 + sin ( Ωk ) ]
The deviation, e , is modeled as white noise and is neglected in (3-7). The notation is
simplified by letting
1 1 2 R LSB 1
K = --- ⋅ ----------------------- = --------------
- ⋅ -----------
- and L = ------------- (3-8)
A G LSB ⋅ R L N
2 –1 R L 1 + K
which gives the expression of the current through the termination load as
L⋅K
I L ( X ) = A ⋅ ( I LSB + G LSB ⋅ V DD ) ⋅ K ⋅ 1 – ------------------------------------- (3-9)
1 + L ⋅ sin ( Ωk )
46
3. DAC Performance Modeling
∑ [ – L sin ( Ωk ) ] n
1
--------------------------------- = =
1 + L sin ( Ωk ) n≥0
When determining the SFDR, the gain and dc level of the signal are neglected, and by
using trigonometric formulae (3-11) becomes
--------------------------------- = f ( Ωk ) – sin ( Ωk ) ⋅ L 1 + ∑ ( L ⁄ 2 ) 2n ⋅
1 2n + 1
–
1 + L sin ( Ωk ) n
n≥1
– cos ( 2Ωk ) ⋅ 2 ∑ ( L ⁄ 2 ) 2n
2n
(3-12)
n – 1
n≥1
where f ( Ωk ) denotes a function containing the dc component and higher frequency com-
ponents that do not affect the SFDR and p = ------------------------ .
p!
q q! ( p – q )!
The SFDR is found as the power ratio between the fundamental and the harmonic as
1 + ∑ ( L ⁄ 2 ) 2n 2n + 1
2
L 2 n 1 + 1 – L2
2
n≥1
SFDR = ----- ------------------------------------------------------------
- = ---------------------------- (3-13)
2n 2n
n∑
4 L
( L ⁄ 2 )
n – 1
≥1
R LSB ⁄ R L 2N – 1
2
SFDR = 1 + 2 ---------------------
- ⋅ 1 + 1 + ---------------------
- (3-14)
2N – 1 R LSB ⁄ R L
When the impedance ratio R LSB ⁄ R L and the number of bits, N , are large, (3-14) is
approximated by
1 R LSB 2
-----------
SFDR ≈ ------------------
- ⋅ - (3-15)
22( N – 2) RL
R LSB
SFDR ≈ 20 log -----------
- – 6( N – 2) (3-16)
RL
This gives a hint on how to choose the output impedance of the LSB current source,
when knowing the specified number of bits, N . The average output resistance, R DC , is
given by
2
R DC = R LSB ⋅ --------------
N
- (3-17)
2 –1
Using (3-17) in (3-16), it is seen that the SFDR also can be expressed as
47
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
2N – 1
R DC ⋅ --------------- R DC
2
SFDR ≈ 20 log ------------------------------- – 6 ( N – 2 ) ≈ 20 log ---------
-+6 (3-18)
RL RL
In Fig. 3-4 we show the simulated and calculated SFDR for a 12-bit DAC. It is seen that
the calculated values fit the simulated curve very well. For high resistance ratios,
R LSB ⁄ R L , the spurious tones are hidden in or below the average quantization noise floor,
equation (1-1), and therefore the simulated curve is saturated, since peaks from noise
components are higher than the spurious tones.
Calculated Value
100
80
SFDR [dB]
60
Simulated Value
40
20
0 2 4 6 8
10 10 10 10
RLSB / RL
Figure 3-4: SFDR vs. resistance ratio when varying the output resistance of a 12-bit DAC.
The important aspect is that if the number of bits should be increased without decreasing
the SFDR, the output impedance for the unit current source has to be approximately dou-
bled for each extra bit of resolution (3-16).
The influence of the output impedance may be minimized by using an operational ampli-
fier at the output which has a very low input impedance, hence the output impedance over
termination impedance is very large [37, 56]. However, in wideband applications the
operational amplifier will limit the performance, due to its limited speed and bandwidth.
48
3. DAC Performance Modeling
error as a current source in parallel with the nominal current source, as shown in Fig. 3-5.
All error sources can be summed and modeled as one error current source connected to
the output of the DAC.
IMSB ILSB
∆IMSB ∆ILSB
The number of error sources connected to the output is determined by the digital input
code, X . Since the mismatch error is considered to be fixed for a certain current source,
the total error is signal-dependent and distortion is introduced.
The matching issue is also further discussed in Sec. 4.1.1 where the current sources used
in the implementation of the Nyquist-rate DAC are discussed.
Suppose in the simplest case that there is an error current, ∆I i ( X ) , in the current source
controlled by the i -th LSB, we have
where δ i ( X ) is the relative current error in the i -th LSB, and x i is the bit controlling the
i -th LSB. If δ i ( X ) = δ i for all inputs, X , (i. e., the size of the error current in a specific
bit position, i , is signal-independent), the total output current is
The error will have the characteristics of a pulsed signal, with different duration and
amplitude, depending on the error position, i . It is seen that mismatch errors in the more
significant bits (when i is large) will affect the signal stronger, due to the 2 i – 1 term. Due
to pulsed error currents in the MSBs, harmonics of odd order occur. The amplitude, A e, i ,
of such an error pulse in the i -th LSB position, as found in (3-21), is given by
A e, i = δ i ⋅ I LSB ⋅ 2 i – 1 (3-22)
49
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
wave with period M and have a duration of M ⁄ 2 (as the MSB current source is switched
to the output). The Fourier transform of the error current in the i -th LSB current source,
F em ( ω, i ) , is approximately given by
∞
2 A e, i 2π
F em ( ω, i ) ≈ ------
jπ ∑ 2k – 1
- ⋅ δ ω – ------ ⋅ ( 2k – 1 )
--------------
M
(3-23)
k = –∞
where M is the signal period and ω is the angular frequency variable, in (3-23) δ [ … ]
describes the Dirac function. The normalized power of the largest harmonic, P k = 2 , (third
order), found by letting k = 2 in (3-23), is given by
1 δ i ⋅ I LSB ⋅ 2
i 2
P k = 2 ≈ --- ⋅ ---------------------------- (3-24)
2 3π
Suppose a sinusoidal input with an amplitude given by (3-6), the normalized signal
power can be written as
2
1 2N – 1
P signal = --- A 2 = --- ⋅ --------------- ⋅ I LSB
1
(3-25)
2 2 2
where A is the amplitude of the signal, and N is the number of bits. The power ratio the
sinusoidal (3-25) and the largest harmonic (3-24) is the SFDR given by
2 N–1 2
1 2N – 1
--- ⋅ --------------- ⋅ I LSB 2---------------
3π 2 i
2
2 2 3π ⋅ (-------------------
2 2 N – i )2
SFDR ≈ ---------------------------------------------- ≈ 2
-----
- ⋅ ----------------------- ≈ -----
- (3-26)
1 δ i ⋅ I LSB ⋅ 2
i 2 δ i2 2 δ i2
--- ⋅ ----------------------------
2 3π
3π
SFDR ≈ 20 log ------ + 6 ( N – i ) – 10 log δ i2 (3-27)
2
Since for square wave signals the third harmonic dominates the other harmonics, (3-27)
is valid for errors in more significant bits. For errors in the least significant bits, the for-
mula is approximately true.
It is seen that for an increase of number of bits, N , the SFDR is improved by approxi-
mately 6 dB, if the increase of the number of bits does not affect the mismatch in the given
bit position. For a one per cent error, δ 14 = 0.01 , in the MSB of a 14-bit DAC, (3-27)
gives SFDR ≈ 53 dB. It can also be seen from (3-27) that if there is mismatch in the MSB
( i = N ) the SFDR is not depending on the number of bits, since the largest mismatch at
the MSB does not decrease with the increase of number of bits. This is also verified in
simulations.
Shown in Fig. 3-6 is the simulation result for a 14-bit DAC. It can be seen that if the rel-
ative error current in the MSB current source is δ 14 = 0.01 , the simulations show
SFDR ≈ 53 dB. For smaller deterministic mismatch errors in the least significant bits, the
50
3. DAC Performance Modeling
error can not be modeled as contribution to the distortions. In this case the spurious tones
are hidden in the noise floor as shown in Fig. 3-6.
110
100
90
SFDR [dBc]
80
70
60
50
40
0
0.5 2
4
1 6
8
1.5 10
12
2 14
Mismatch (%)
Figure 3-6: Deterministic variations of mismatch over bit positions in a 14-bit DAC. SFDR is shown
as a function of the mismatch error bit position and size. For small errors the spurious
tones are hidden in the noise floor.
sources are connected in parallel. Suppose all mismatch errors in each LSB current source
are uncorrelated. Then, the mismatch error in the i -th LSB current source, δ i , has the
mean value
2i – 1 2i – 1
µi = E ∑ δk = ∑ E { δk } = 0 (3-28)
k = 1 k=1
One of the main issues is actually to find out how the correlation between two error
sources affects the result. Since the current sources generally are laid out close to each
other, the matching errors can be highly correlated. The mean value of the error sources
can in any case be neglected since its influence only give rise to a linear gain or offset
51
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
error.
The total error current due to the stochastic mismatch can be summed into one source
and the expectation value of the total generated error power can be expressed by using the
equations from the previous section. If it is assumed that the third harmonic dominates
other harmonics, with (3-24) and (3-25), the expectation value of the SFDR can be ana-
lyzed. Due to the summation of 2 i – 1 stochastic sources, the amplitude, A i , in the i -th
LSB, for the third harmonic as was seen in (3-23) is
2
A i = ------ ⋅ δ LSB ⋅ I LSB ⋅ 2 i – 1 (3-30)
3π
The normalized signal power is given by (3-25) and the normalized power of the third
harmonic, E { P k = 2, tot } , is derived, by summing the error current amplitudes as given in
(3-30)
2
N
N
δ LSB ⋅ I LSB ⋅ 2 i – 1
E { P k = 2, tot } = E ∑ x i ⋅ A i = E ∑ x i ⋅ 2 ---------------------------------------------- =
i = 1 3π
i=1
N 2
I LSB 2
= ------------- ⋅ E ∑ xi ⋅ 2i – 1 ⋅ δ LSB (3-31)
3π ⁄ 2
i=1
Noting that x i = x i2 since x i ∈ { 0, 1 } and that the error sources are uncorrelated, the
expression in (3-31) is rewritten as
N
I LSB 2
E { P k = 2, tot } = ------------- E δ LSB ⋅ ∑ x i2 ⋅ 2 i – 1 =
2
3π ⁄ 2
i=1
I LSB 2
= ------------- E { δ LSB
2 ⋅ ( x 1 + 2x 2 + … + 2 N – 1 ⋅ x N ) } =
3π ⁄ 2
I LSB 2
= σ LSB
2 ⋅ ------------- ⋅ E { X } (3-32)
3π ⁄ 2
where the digital input signal, X , was defined in (2-2). The mean value, E { X } , is the dc
value of the sinusoidal and given by
2N – 1
E { X } = --------------- (3-33)
2
Combining (3-32) and (3-33), we have
I LSB 2 2N – 1
E { P k = 2, tot } = ------------- ⋅ σ LSB
2 ⋅ --------------- (3-34)
3π ⁄ 2 2
The expectation value of the SFDR can be derived by combining (3-25) and (3-34)
52
3. DAC Performance Modeling
2
1 2N – 1
--- ⋅ --------------- ⋅ I LSB
2 2 3π
2 2N
⋅ -----------
E { SFDR } = ----------------------------------------------------------
- ≈ -----
- - (3-35)
I 2 N–1 4 σ 2
------------- ⋅ σ 2 ⋅ ---------------
LSB 2 LSB
3π ⁄ 2 LSB 2
In dB, this becomes
3π
E { SFDR } ≈ 20 log ------ + 3N – 10 log σ LSB
2 (3-36)
4
SFDR dist. Mismatch dev.: 1/128 SFDR dist. Mismatch dev.: 1/512
40 45
Number of occurences
35
30
30
25
25
20
20
15
15
10
10
5 5
0 0
95 100 105 110 115 120 125 130 110 115 120 125 130 135 140 145
SFDR [dBc] SFDR [dBc]
SFDR dist. Mismatch dev.: 1/1024 SFDR dist. Mismatch dev.: 1/2048
35 45
Number of occurences
35
25
30
20
25
20
15
15
10
10
5
5
0 0
115 120 125 130 135 140 145 120 125 130 135 140 145
SFDR [dBc] SFDR [dBc]
Figure 3-7: Histograms showing the variation of the SFDR vs. mismatch for a 12-bit DAC. a) σLSB =
1/128 b) σLSB = 1/512 c) σLSB = 1/1024 d) σLSB = 1/2056
Histograms, shown in Fig. 3-7, using 1000 FFTs of the output of a 12-bit DAC are sim-
ulated to find the expectation value of the SFDR. Each FFT uses 57643 points. The values
are given for different mismatch error standard deviations (which are rather small com-
pared to a real process [4, 9]). The different SFDR expectation values found are
a) σ LSB = 1 ⁄ 128 gives E { SFDR } ≈ 110 dB
b) σ LSB = 1 ⁄ 512 gives E { SFDR } ≈ 122 dB
c) σ LSB = 1 ⁄ 1024 gives E { SFDR } ≈ 128 dB
d) σ LSB = 1 ⁄ 2048 gives E { SFDR } ≈ 134 dB
53
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
It is seen that the SFDR is decreased with approximately 6dB for every time the standard
deviation is doubled. In the figures the yield of the circuits is shown as well. The 90%
yield implies that 90 per cent of the chips will have a better SFDR than the value shown
in the figures.
In Fig. 3-8 we see the comparison between the calculated values from (3-36) and simu-
lated values from histograms as shown in Fig. 3-7. The curves fit well together. For low
values of error standard deviation the spurious tones are hidden in the noise floor and
therefore the curve is saturated. The simulation results from a 20-bit DAC are also
reported in [76].
105
Calculated Value
100
95
90
Simulated Value
SFDR [dB]
85
80
75
70
65
60
55
−4 −3 −2 −1 0
10 10 10 10 10
Deviation in the LSB current
Figure 3-8: Summary of the histograms for a 12-bit DAC compared with the calculated values.
Expectation value of SFDR vs. standard deviation for the mismatch error. For low errors
the simulated spurious tones are hidden in the noise floor.
To the matching error we can also include influence of voltage drops over internal wires
as well as the influence of non-linear switches, etc. This also affects the value of the cur-
rent fed to the output, hence can be modeled as matching errors.
54
3. DAC Performance Modeling
ILSB
ilsb
Figure 3-9: Nominal unit current source, ILSB, with included noise source, ilsb.
With each unit current source (equal to the LSB current source) a normal distributed
noise source, i lsb , is associated. Within a certain bandwidth, BW , the mean value is
2
E { i lsb } = 0 and its variance is given by E { i lsb } = i lsb
2 . The expectation value of the
2 = E 2
P noise = i tot ∑ i lsb, k = i lsb
2 ⋅X (3-38)
k
where X is the mean value of the digital input, hence the number of sources that are in
mean connected to the output. Assuming that X is a full-scale sinusoidal, the mean value
is given by
2N – 1
E { X } = X = --------------- (3-39)
2
where N is the number of bits. Combining (3-38) and (3-39) gives the expression of the
total normalized noise power, P noise , as
N–1
2 = 2
P noise = i tot --------------- ⋅ i lsb
2 (3-40)
2
The normalized signal power for the full-scale sinusoidal is given by
1 2N – 1 2 2
P signal = --- ⋅ --------------- ⋅ I LSB (3-41)
2 2
Combining (3-40) and (3-41), the SNR determined by the noise is expressed as
1 2 N – 1 2 2
--- ⋅ --------------- ⋅ I LSB
P signal P signal 2 2 2 N I LSB
2
SNR = -------------- = -------------- = ----------------------------------------------
- ≈ -----
- ⋅ ---------- (3-42)
P noise 2 2N – 1 2 4 i2
i tot --------------- ⋅ i lsb lsb
2
The approximation is valid when the number of bits, N , is large. The SNR in dB is
For wideband CMOS DACs it is the thermal noise from the channel resistance that dom-
55
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
inates other noise sources [37]. The power spectral density (PSD) of a thermal noise cur-
rent source as modeled in Fig. 3-9, is described by
i n2 8
------ ≈ --- kT g m (3-44)
∆f 3
where k is the Boltzmann’s constant, T is the absolute temperature, and g m is the small-
signal transconductance of the current source transistor. Equation (3-44) holds for both
single transistor current sources and cascoded current sources if the impedance at the
source of the cascode transistors is very low. The unit current is approximately given by
β
I LSB = --- ( v SG – v T ) 2 (3-45)
2
g m ≈ 2β ⋅ I LSB (3-46)
Using the noise and transconductance definitions in (3-44) and (3-46) and assuming that
the thermal noise in the transistors dominates, the noise spectral density in the unit current
source is
8
i n2 ⁄ ∆f ≈ --- kT 2β ⋅ I LSB (3-47)
3
If we assume a certain noise bandwidth of BW , the normalized total output noise power,
2 , from the unit current source is
i lsb
2 = 8
i lsb --- kT ⋅ BW ⋅ 2β ⋅ I LSB (3-48)
3
All the thermal noise associated with the switched current sources degrades the SNR.
Substituting the values from (3-48) into (3-43) gives the approximate expression
In Fig. 3-10 the simulated and calculated SNR for a 12-bit and a 14-bit DAC vs. the cur-
rent through the unit current source, I LSB , are shown. In the simulations process data from
Ericsson’s 0.6µm CMOS process were used. It can be seen from Fig. 3-10 that for high
unit currents, the thermal noise is lower than the quantization noise. The simulated curves
fit well to the calculated values.
The formulae in this section give a guidance on how to choose the current of the unit
(LSB) current source.
56
3. DAC Performance Modeling
Simulated and calculated SNR. 12−bit DAC. Simulated and calculated SNR. 14−bit DAC.
100 120
Sim
Calc
100
80
80
SNR [dB]
SNR [dB]
60
60
40
40
20
20
0 0
a) −12
10 10
−10 −8
10 10
−6 b) −12
10
−10
10 10
−8 −6
10 10
ILSB [A] ILSB [A]
Figure 3-10: Simulated and calculated SNR vs. the ILSB current for a) 12-bit and b) 14-bit DAC.
Cout(X)
RL
iL(t)
Figure 3-11: Generalized view of the converter with output capacitance and resistance.
When the input switches from one code X i – 1 to a consecutive code X i = X i – 1 + 1 , the
output will ideally be given by a step function. However, due to reactive parts of the out-
put and load impedance, the signal will settle towards an end value. In Fig. 3-11 the first
order model of the output of the DAC is shown. In this brief discussion, the influence of
a load capacitance is neglected.
The output current, i L ( t ) , through the termination resistance, R L , is described by
V DD – R L ⋅ i L ( t ) d
i L ( t ) = I S ( X ) + --------------------------------------
- – C out ( X ) ⋅ R L ⋅ ----- i L ( t ) (3-50)
R out ( X ) dt
57
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
i L ( t ) = I i – 1 + ( I i – I i – 1 ) ⋅ ( 1 – e –t ⁄ τi ) (3-51)
where
I LSB + V DD ⋅ G LSB
I i = --------------------------------------------- ⋅ X i (3-52)
1 + R L ⋅ G LSB ⋅ X i
as from (3-2) with the assumption that the start value also has settled to its final value. The
time constant is given by
R L ⋅ C LSB
τ i = ------------------------------------------- ⋅ X i (3-53)
1 + R L ⋅ G LSB ⋅ X i
where C LSB is the capacitor associated with the LSB current source. Obviously τ i is sig-
nal-dependent, and therefore the slewing is non-linear. The maximum value of the time
constant is found when X i is maximal, hence when X i = 2 N – 1 . For the maximum val-
ues at the output the slewing is more crucial, yielding a lower bandwidth. This is also
found in measurements, since non-full-scale signals can give rise to a better SNR and
SFDR than a full-scale input signal. The result in (3-53) may be used to find out the max-
imum possible sampling frequency as well.
If the converter is well designed with respect to the output impedance, (3-53) may be
approximated by
τ i ≈ R L ⋅ C LSB ⋅ X i (3-54)
The knowledge about the time-constant gives the designer information how to construct
the current sources in order to reach the specified update frequency. The error in ampli-
tude level is determined by the time the signal is allowed to settle.
58
4. A CMOS DAC Chipset for VDSL Applications
The first generation focuses on a low-voltage converter and is designed by N. Tan [64].
The second generation focuses on a wideband version and the third generation focuses on
improved matching for the wideband applications [68]. The design of the second genera-
tion converters is presented in this chapter and the target was to design a 14-bit converter
with low power consumption, high-resolution, and high bandwidth. For the second and
third generation, also the knowledge from the modeling work as described in Chapter 3
has been applied. We have however found out that matching was not considered properly
for the second generation and this will also have a large impact on the achieved results
(Sec. 4.3), especially the linearity of the converter.
The first and second generation converters have been measured and the third generation
is still to be measured. The structure used for all converters in the chipset is the current-
steering DAC as discussed in Sec. 2.2.
59
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
• The digital part consisting of clock, flip-flops stabilizing the digital input, and an
MSB segmentation part.
• The analog part consisting of an array of unit current sources.
• The mixed analog/digital parts consisting of current switches.
These issues are further discussed in the following subsections.
Unit
Current
Mixed Digital/Analog Parts
Sources
I- I+
Dummy Sources
Figure 4-1: Overall structure of the converters in the chipset.
For the i -th LSB, 2 i – 1 unit current sources from the array are connected together and
their output is fed to the current switch controlled by the i -th bit. The output is a differ-
ential current output and if the corresponding bit is a “1” or “0”, the current is routed to
the positive, I + , or the negative output, I - . Differential signals will decrease the influence
of even harmonics (due to non-linearity) in the output.
The segmentation of the MSBs introduces a delay for the MSBs and to avoid skew
between the bits, a digital delay line is used for the LSBs. A clock tree is implemented to
get proper clock delay to all flip-flops controlling the current switches.
60
4. A CMOS DAC Chipset for VDSL Applications
2β 2 ( 1 + η ) 2β 3
R I ≈ ( 1 + η ) ⋅ ------------------- and R ≈ R ⋅
I ------------------------------- (4-1)
λ 2 ⋅ I src 3/2 II
λ ⋅ I src
where I src is the output current of the source, λ is the channel length modulation (ap-
proximately equal for all the transistors), β i is the transconductance parameter of transis-
tor i , and η is a parameter given by the bulk-source connection of the transistor (Sec.
1.4.1 in Part A: Overview: Digital-to-Analog Converter Design).
VDD
VDD
M1
M1
M2
M2
Isrc M3
a) b) Isrc
Figure 4-2: Schematic view of PMOS current sources with a) single and b) double cascode.
According to (4-1), the output impedance of the double cascode is higher than for the
single cascode. However, there is one additional internal node, which increases the capac-
itive part of the output impedance and hence introducing extra poles and longer settling
time and therefore this constellation may degrade the performance.
a) b)
Gates M1
M1
M2
M2
M3
Outputs
Figure 4-3: Layout view of PMOS current sources with a) single and b) double cascode.
The transistors all have the same width, W = 2µm , and the lengths are given by
L 1 = 8µm , L 2 = 2µm and L 3 = 1.2µm , for the transistors M1, M2, and M3, respec-
tively.
In the 14-bit DAC the binary weighted current sources were constructed by using unit
current sources from the array by starting in one corner and use one unit current source
for the LSB, the next two attaching unit sources for the second LSB, etc. This will how-
ever not give good matching properties and improvements will be discussed in Chapter 5.
61
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
A part of the second generation array containing the unit current sources is given in Fig.
4-4a). In the figure the internal nodes of the current sources are connected together,
increasing the output capacitance of the sources. The V DD connections are routed over
the M1 transistors, which influences the matching of the source transistors [73]. This were
changed in the third generation DAC as shown in Fig. 4-4b). The internal connections
were also removed to reduce the capacitance.
a) b)
VDD Connections
moved away
VDD Connections from M1
over M1 transistors
M1
Connected Unconnected
Internal M2 Internal
Nodes Nodes
M3
Output Output
Figure 4-4: Part of current source array for the a) second and b) third generation DAC with double
cascode current sources.
For a 14-bit DAC with the peak output current of 20mA, the output current for the unit
current source is
–3
20 ×10
- ≈ 1.22µA
I unit = ------------------- (4-2)
2 14 – 1
Consider the approximate drain current in saturation region for the PMOS transistor
β
i D ≈ --- ⋅ ( v SG – v T ) 2 (4-3)
2
There are several components that are affected by matching errors, i.e., transistor size,
threshold voltage, supply voltage, oxide thickness, bias voltage, output voltage, etc. By
differentiating (4-3) the influence of the matching errors can be seen as
∆i D ∆β ∆V SG – ∆V T
- = ------- + 2 -------------------------------
-------- (4-4)
iD β V SG – V T
When assuming that the deviation of the source-gate voltage is very small, (4-4) approx-
imately becomes
∆i D ∆β ∆V T
- ≈ ------- – 2 ---------------------------
-------- - (4-5)
iD β ( V SG – V T )
The β and V T matching are given by the following terms from (4-5)
∆β 2∆V T
------- and ----------------------
- , respectively (4-6)
β V SG – V T
62
4. A CMOS DAC Chipset for VDSL Applications
From studies in the literature [4, 9, 55, 71], we know that the β and V T matching are
uncorrelated and that their variance is inversely proportional to the transistor area, WL
Since low power consumption and saturated transistors are wanted, V SG – V T is kept
low and therefore the expression (4-6) may be large.
From a geometrical point of view the term ∆β ⁄ β in (4-4) are dependent on the layout
style. Large transistors and special layout styles are used to reduce the influence of these
kinds of mismatch, i.e., common-centroid and interdigitized layout styles. In the current
source array the matching errors between two neighboring unit current sources may be
strongly correlated and therefore the binary weighted current sources should actually be
constructed by using unit current sources from different positions in the array [9, 39]. This
may however increase wire resistance and capacitance.
If the SFDR of the converter should be guaranteed to be higher than for example 80dB,
equation (3-16) states that
6 ( N – 2 ) + 80
----------------------------------
R LSB ≥ R L ⋅ 10 20 (4-7)
With a 50Ω termination the bound on the output impedance of the current source is
found to be
R LSB ≈ 2GΩ (4-8)
which is hard to accomplish. The output impedance is also dependent on the bias current
controlling the source-gate voltage of the current source (4-1). These voltages are simply
generated using a cascoded current mirror and the reference current through the primary
side of the mirror is determined by a terminating resistance (at an approximate size of 3-
5kΩ). The cascoded current source has an (simulated) output impedance of approximately
R out ≈ 600MΩ (4-9)
corresponding to an SFDR of 70dB for the 14-bit version, which is sufficient for the want-
ed applications.
63
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
G sw ≈ β ⋅ ( V G – V T – V D ) (4-10)
VDD
VDD
Ii = 2i-1 Iunit
Ii = 2i-1 Iunit
VC
VC
φ φ
φ φ
I+ I- I+ I-
Figure 4-5: Differential current switch as a) circuit model and b) transistor implementation.
In principle, to keep the voltage drop over all switches and current sources nearly con-
stant, hence proper matching simpler, the voltage node V C must be constant and equal
for all bits. For the i -th LSB we let this node voltage be denoted V C, i and it is approxi-
mately given by
where I unit is the output current of one unit current source, V DC is the average voltage
at the output, G sw, i is the conductance of the switches, and G unit is the output conduc-
tance of the unit current source. Equation (4-11) is rewritten as
G sw, i
I unit + G unit ⋅ V DD + ------------ - ⋅ V DC
2i – 1
V C, i ≈ -------------------------------------------------------------------------------- (4-12)
G sw, i
G unit + ------------ -
2i – 1
In order to keep the V C, i equal for all bits, G sw, i ⁄ 2 i – 1 must be equal for all bits, hence
where G sw, unit is the conductance of a certain unit switch, i.e., the switch corresponding
to the LSB. This implies that the sizes of the switches must be chosen so that their con-
ductance is increasing exponentially with higher LSBs. This will however give very large
switches for the MSBs and too large clock load as well as too high CFT due to the larger
Miller capacitance. A trade off has to be done.
64
4. A CMOS DAC Chipset for VDSL Applications
The layout of the current switch for the LSB is shown in Fig. 4-6, where one of eight
transistors in parallel is used as switch for each channel. To achieve equal capacitive load
for all switches, the number of NMOS transistors (eight) in the switches is the same for
all bits, although some of the transistors are shorted for LSBs, as shown in Fig. 4-6. A
shielding ground plane around all switches is also used.
Shielding
Ground
Positive switch input Positive Output
Current Input
The inverse switch signal is generated with a simple CMOS inverter. A better glitch per-
formance can be found with a more proper switching scheme. When both switches are
turned off at the same time, the voltage at the output of the current source, V C, i , will
increase and the glitch will be larger when one of the switches is turned on again. There-
fore, it should be guaranteed that both switches are not completely off at the same time
[9, 71], as sketched in Fig. 4-7.
Switch
Ampl.
Positive Switch Signal
VDD
The possible switching voltage swing, V s, max , of the switches is determined by the max-
imum threshold voltage of the NMOS transistors, V T , max , and the maximum output volt-
age, V out, max , since
V s, max = V out, max + V gs, switch > V out, max + V T , max (4-14)
For a standard process with a maximum output current of 20mA terminated with a 50Ω
load, the minimum switching voltage is approximately V s, max ≈ 1.7 V. For the low-volt-
age ( V DD = 1.5 V) versions of the DACs a charge-pump circuit is used to increase the
voltage swing of the switching signals to overcome these problems.
65
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
[ φ a, φ b, φ c ] = [ φ 1 ∧ φ 2, φ 1, φ 1 ∨ φ 2 ] (4-15)
Using the thermometer code we can use a number of equally large current sources, Fig.
4-8. The layout can be more regular and current source sizes as well as current switch
sizes are equally large for the segmented bits. The digital circuitry that transforms the
binary code into the thermometer code introduces delay and therefore the LSBs have to
be delayed correspondingly long time as shown in Fig. 4-1.
I 2I I I I
φ1 φ2 φa φb φc
Figure 4-8: Example of segmentation of two binary weighted sources into three equally large sources.
The current that will flow through one switch will be smaller than for the MSB without
segmentation and therefore a smaller voltage drop over the switches which further
reduces the distortion.
When segmenting the four MSBs in the 14-bit DAC, they will turn into 15 equally large
current sources, each one containing 1024 unit current sources. This is also a disadvantage
since the number of switches, the extra introduced logic, and wires are exponentially
growing with number of bits to segment.
The glitches are roughly dependent on how many bits that are switching between two
consecutive values at the input. By setting up a cost function, i.e., the power of the glitches
or the number of switching bits, we can examine how the power of the glitches are varying
with the number of segmented bits. Simulation results for a 14-bit DAC, shown in Fig. 4-
9, state that about 4-5 segmented bits give a large improvement. For more segmented bits,
the improvement is not that high, and more digital circuitry is also introduced, hence
higher complexity and power consumption.
The segmentation guarantees monotonicity (for the MSBs) and improved DNL since for
an increasing input additional current sources are connected to the output. When only
some of the bits are segmented, the INL is not improved. Matching of the segmented bits
may further be improved by using randomization, averaging techniques, or calibration, as
described in Chapter 5.
66
4. A CMOS DAC Chipset for VDSL Applications
53
52
Power [dB]
51
50
49
48
47
46
1 2 3 4 5 6 7 8 9 10
Number of segmented bits
Figure 4-9: Power of glitches as a function of the number of segmented bits in a 14-bit DAC.
67
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
VDD R V1 R V2 R R VN
IDD
Figure 4-10: Model of the voltage supply line connected to a number of current sources.
In high-performance circuits the reference current errors have to be very small. There-
fore, we cannot distribute the supply along a single wire due to the voltage drop. For the
supply distribution it is also necessary to analyze the peak current to avoid breakdown of
the wires.
Digital Inputs
Digital
Circuitry
Interconnection
Wires
Current
Source Array
Analog Inputs
Current Outputs
Figure 4-11: Chip layout of the third generation DAC.
In mixed analog/digital design, disturbances from the digital to the analog part (or vice
versa) spread along supply lines and the substrate. The substrate coupling may be strong.
It is therefore necessary to do careful designs with proper shielding, which may be done
in different ways [33], i.e., guard rings, grounding, etc.
At board level it is also needed to shield and separate the signals [20]. The chip analog
pins are separated from the digital as far as possible. As well as grounding pins are used
in-between. The pin configuration of the chip is equal to the Analog Devices’ 14-bit DAC:
AD9764 (member of the TxDAC series) and hence a 28-bit SOIC socket [84].
A good clock distribution is needed to minimize the influence of clock skew, which
strongly affects the performance of the converter. The energy of the glitches are depen-
dent on the skew between the different bits. Shown in Fig. 4-12 are two examples on good
clock distribution that reduce the skew between different subcircuits.
68
4. A CMOS DAC Chipset for VDSL Applications
Subcircuits
Subcircuits
Master Master
Figure 4-12: Clock tree structures to improve buffering and equal clock delay to the circuit.
a) b) ADC DAC
DAC
Test Ref Test
c) d)
DAC DAC
Test Test
DAC ADC
Ref Ref
Figure 4-13: General measurement setups. a) Input is generated with a stimuli generator b) Input signal
is generated using an ADC c) Output is compared with a reference DAC d) The output is
converted back to an analog signal and can be compared with the original input signal.
At the behavioral-level, MatLab or similar tools are used (Chapter 3), on the circuit-
level, Spice, Spectre or similar [43] are used. Since DACs are mixed analog/digital cir-
cuits mixed-mode simulations are performed. It may be necessary to partition the circuit
into digital and analog parts to increase the speed of the simulation [43]. The digital cir-
cuits may be simulated with event driven simulators, using simplified models for transis-
69
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
tors. However, at the interface between the digital parts and analog parts the analog
signals may be affected by digital switches, e.g., current switches driven by digital switch-
ing signals. In order to analyze glitches, supply noise, substrate coupling, etc., it may be
preferred to perform the whole simulation with a more accurate analog simulator using
continuous-time models.
We will also divide the measurements into single-, dual-, and multi-tone measurements
since all necessary information about the converter may not be extracted from single-tone
measurements [5].
a) 80 b) 80
60 60
40 40
Magnitude [dB]
Magnitude [dB]
20 20
0 0
−20 −20
−40 −40
−60 −60
A signal that is not completely periodic will also give rise to distortion in the output. In
Fig. 4-15b) the amplitude spectrum of the output signal from a 14-bit DAC when leaving
out only one single sample of the whole period containing 1917 points is shown.
70
4. A CMOS DAC Chipset for VDSL Applications
80 80
a) b)
60 60
40 40
Magnitude [dB] 20 20
Magnitude [dB]
0 0
−20 −20
−40 −40
−60 −60
−80 −80
−100 −100
For the discrete multi-tone (DMT) we use a signal equal to orthogonal frequency divi-
sion multiplexing (OFDM) where the signal, y ( t ) , should be given by
N – N0 – 1
1
y ( t ) = --------
N
∑ c n ⋅ e jn∆ωt (4-16)
n = N0
where c n is the complex value given by the code word assigned to the n -th carrier. ∆ω
is the spacing between the carriers, N 0 ⋅ ∆ω is the starting frequency, and N is the num-
ber of carriers.
When simulating the DMT signal, the first approach would be to simply use the expres-
sion where no information has been applied to the c n terms from (4-16)
K
y(t ) = A ⋅ ∑ sin ( kω0 t ) (4-17)
k=1
71
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
Signal amplitude.
50
10
0 0
−10
−50
−20
−100
−30
−150
−40
−200 −50
0 0.02 0.04 0.06 0 0.02 0.04 0.06
Time. Time.
Figure 4-16: The output signal for a DMT system with a) fixed phase and b) random phase. The crest
factor is improved by changing to random phase. Note the different scales.
These are some of the aspects that have to be considered when analyzing the output sig-
nal. With an erroneous input signal the output cannot be trusted.
72
4. A CMOS DAC Chipset for VDSL Applications
a) b)
Figure 4-17: a) 10-bit single cascode DAC. 3.3 V supply. Single tone. 20 MHz clock frequency.
b) 14-bit double cascode DAC. 5 V supply. Dual tone. 20 MHz clock frequency.
The SFDR vs. clock frequency for the single cascode 10-bit DAC, single cascode 14-bit
DAC, and double cascode 14-bit DAC is shown in Fig. 4-18a) through c), respectively.
In the figures the results from both 3.3V and 5V supplies are shown.
73
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
50 (49)
48
SFDR [dB]
46
(44)
44
42
40
0 5 10 15 20 25 30 35 40 45 50 55
Clock Frequency [MHz]
SFDR vs. Clock Freq. 14−bit DAC (Single Casc.) 3.3 and 5−V supply.
50
(49)
48
SFDR [dB]
46
44
(42)
42
40
0 5 10 15 20 25 30 35 40 45 50 55
Clock Frequency [MHz]
SFDR vs. Clock Freq. 14−bit DAC (Double Casc.) 3.3 and 5−V supply.
51
50
(49)
49
48
47
SFDR [dB]
46
45
44
43 (42)
42
41
40
0 5 10 15 20 25 30 35 40 45 50 55
Clock Frequency [MHz]
Figure 4-18: SFDR vs. Clock frequency for the a) single cascode 10-bit DAC, b) single cascode 14-bit
DAC, and c) double cascode 14-bit DAC. (*) at 3.3 V supply. (o) at 5 V supply. The
signal-to-clock frequency ratio is approximately 1/6.
74
5. Improvements of the DAC Design
5 Improvements of
the DAC Design
It has been shown that the performance of the converter is dependent on matching, noise,
and output impedance. In measurements (Chapter 4) it has also been found that it is the
non-linearity due to matching errors that limits the performance of the implemented con-
verter chipset. In this chapter we present some techniques that can be used to improve the
performance of the wideband converters. Some of the techniques will be implemented in
future versions of the chipset.
The most common techniques to further improve the performance are
• Digital correction; calibration and trimming of sources, etc. [8, 10, 13, 54, 56]
• Dithering; randomization of segmented sources and dynamic element matching,
etc. [11, 24, 35, 56]
In some cases, a kind of an observer is needed. For the use in DACs it has to be analog
and it is therefore often hard to design due to its use in wideband applications.
With fine quantization, hence a large number of bits, the quantization error is considered
to be white noise and can therefore be neglected in this discussion. Suppose that the con-
verter is non-linear of the first order and that the output amplitude can be written as
75
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
Note that no delay skew is assumed in this simple model. Suppose that a new input, X̃ ,
which is described by
X̃ = X + b ⋅ X 2 (5-3)
is applied to the input. By inserting (5-3) in (5-2), the output signal becomes
It is obvious that even more distortion components appear in the output. However, if
A ref b + a = 0 ⇒ b = – a ⁄ A ref (5-5)
a2 a3 4
A out ( X ) = A ref ⋅ X – 2 ---------- X 3 + ---------
2
-X (5-6)
A ref A ref
If it is guaranteed that
a2 a3 4
a ⋅ X 2peak > 2 ---------- X 3peak and a ⋅ X 2peak > ---------
2
- X peak (5-7)
A ref A ref
the SFDR is improved since the higher order harmonics are smaller than the first order
harmonic as the example shows in Fig. 5-1, where the improvement in SFDR is approxi-
mately 27dB with pre-distortion.
0 0
−20 −20
Magnitude [dB]
Magnitude [dB]
∆SFDR
−40 −40
−60 −60
−80 −80
−100 −100
Note that for a sinusoidal input, we may have components that are folded back onto the
fundamental frequency, X 3, X 4, … , and interference occurs.
76
5. Improvements of the DAC Design
Intentional pre-distortion is most often only suitable for measurement, where it is rela-
tively easy to distort the input signal. In a real chip implementation, digital squaring cir-
cuits are needed (5-3), as well as analog observers to determine the magnitude of the
actual non-linearity. This introduces more complexity to the design of the circuit, higher
power consumption, etc.
Original DAC
Outputs have to
be observed.
Logic Current Sources
Compensating DAC
Figure 5-2: Use of extra current sources to compensate and calibrate errors.
This method will however require an observer that performs an analog-to-digital conver-
sion, which is undesired. An automatic compensating routine that works at the same speed
as the original converter is also needed. This implies that the observer must know the
input signal and hence it may be necessary to generate special calibration vectors inter-
nally. This further slows down the converter and increases the complexity. It should also
be noted that the compensating DAC cannot only be set for static errors, since it is the
dynamic errors that determine the quality at higher frequencies.
Calibration or trimming techniques can be used to minimize the influence of mismatch
[8]. The basic principle is that the source-gate voltage, V SG , can be held by using the
source-gate capacitance, C sg , see Fig. 5-3. When switch S 1 and S 2 are conducting (in
the positions opposite to the ones shown in the figure), the capacitance, C sg , is charged.
The V SG will now take a value corresponding to the reference current, I ref , which is the
wanted nominal output current of the current source. Mismatch errors of the transistor will
not affect the output current since V SG is dependent on the reference current only. When
S 1 and S 2 are not conducting the V SG is set by the charge on C sg . However, due to the
leakage currents and CFT the true value cannot be guaranteed, the source-gate voltage has
to be refreshed at regular basis.
The advantage of this technique is that the matching errors of the transistors no longer
affect the result. Disadvantages are that additional switches are needed (hence larger load)
and for high resolutions a large number of current sources have to be calibrated. This also
implies that the calibration phase may take a long time, which may not be possible for cer-
77
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
tain applications. Still we also have to guarantee that the errors in switches, leakage cur-
rents, etc., are smaller than the original matching errors.
Cgs
Vbias S1
S2
Iref
To current
switches
Figure 5-3: Circuit solution to calibrate the current sources at certain time points.
The technique can be used for the segmented MSBs, but than their mismatch with the
binary weighted LSBs may be even worse, hence the reference current source must be
very accurate. The technique is also suitable for R-2R ladder structures where there is only
a number of equally large current sources [8].
5.2 Dithering
The main principle of dithering technique is to smooth distortion into noise or generally
forcing matching errors (or similar) that generate non-linearity to be signal-independent.
Binary Therm.
Input Code
Current
Segmentation Switches
Randomizer
Figure 5-4: Randomization of bits. The matching error becomes uncorrelated with the signal.
The matching errors become more or less uncorrelated with the signal and therefore, the
78
5. Improvements of the DAC Design
errors turn into noise instead of distortion. Assume that there are K segmented current
sources. Each current source, I k , k = 1…K , has the same nominal size, I src . For each
source a relative mismatch error, δ k , is associated, hence
I k = I src ⋅ ( 1 + δ k ) (5-8)
Without the randomization the relative error for the code X M = M = 0011…11 ( M
number of “1”:s) would be
M
1
δ Xi = -----
M ∑ δk (5-9)
k=1
δ1 + δ2 δ1 + δ2 + δ3 δ1 + … + δK
δ X 1 = δ 1, δ X 2 = ----------------
-, δ X 3 = ----------------------------
-, …, δ XK = -----------------------------
-
2 3 K
The DNL and INL values are given by
D i = δ i and I i = ∑ Di = ∑ δi (5-10)
With randomization we have the same relative error for all codes X i as
K
1
δX = ----
K ∑ δk (5-11)
k=1
since for each code all current sources are used equally many times (in mean) due to the
random assignment of sources. Since all codes give the same error, this will only be seen
as an offset error for all the segmented current sources
1
D 0 = ---- ( δ 1 + … + δ K ) and D i = 0 for i ≥ 1 (5-12)
K
and hence the average DNL will be improved. The DNL error at the gap between the seg-
mented MSB sources and the LSBs may increase. The INL values become
1
Ii = ∑ Di = D 0 = ---- ( δ 1 + … + δ K )
K
(5-13)
In the frequency domain the distortion terms are reduced to the cost of a higher noise
floor, Fig. 5-5. The increase of noise is dependent on the matching error sizes, δ k . In the
simulation a mismatch error with standard deviation of approximately 12% was used.
For clock driven assignment of the sources, the errors may still be slightly signal-depen-
dent, but the linearity errors are reduced. It should be noted that these results are very hard
to predict and derive. The clock driven (cyclic shift) has the advantage of not needing any
random sequence. The output signal using this technique will still contain distortion
terms, they are however reduced to the cost of a higher noise floor.
79
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications
0 0
−10 −10
−20 −20
−30 −30
Magnitude [dB]
Magnitude [dB]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
−100 −100
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
Normalized frequency Normalized frequency
Figure 5-5: Output amplitude spectrum from 256 segmented sources a) without randomization and b)
with randomization.
In fact, the principle of randomization can be generalized to be used on the entire DAC.
The current sources in a current source array are addressed basically in the same way as
a RAM. To achieve a simple and regular layout, this forces digital circuits to be included
with the current source in the array, since each current source has to have its own current
switch. This also increases the wire lengths and noise between digital and analog parts.
For DAC structures with resolutions higher than 8-10 bits this technique is hard to apply
due to area and complexity. Hybrid versions use this digital control for the MSBs and a
binary weighted structure for the LSBs, very similar to randomization of segmented
sources.
Another technique is to use dynamic element matching (DEM) where the main principle
is to divide the converter into a number of subconverters with a lower resolution. The
input to the subconverters are modified so that the matching errors (noise) is shaped as
well [24, 35]. Using the randomization technique, the noise floor is increased (see Fig. 5-
5b). To reduce the noise floor within a certain signal band, the noise can be shaped and be
moved to other (higher) frequencies. This higher noise is filtered out by an analog filter
at the output. However, this implies an oversampling converter structure, which is
described in the next part of the thesis, Part C: Oversampling Digital-to-Analog Convert-
ers for ADSL Applications.
80
Oversampling
C
Part C:
Digital-to-Analog
Converters for
ADSL Applications
1 Introduction
In this part of the thesis we present an overall description of the oversampling digital-to-
analog converter (OSDAC) with theory and implementation. First we give a general intro-
duction to the principle of oversampling and in Chapter 2 a description of building blocks
associated with an OSDAC. The design of a 3.3V-5V CMOS OSDAC for ADSL appli-
cations is described in Chapter 3.
A Nyquist-rate converter has the peak SNR (within the Nyquist bandwidth), f N , given
by SNR ≈ 6.02N + 1.76 dB. The sampling frequency, f OSR , for the OSDAC is given by
the oversampling ratio (OSR) and the original sampling frequency, f s , as
The quantization noise will now be given by the bandwidth determined by f OSR . With
a lowpass filter, Fig. 1-1, the signal within the wanted original bandwidth, f N = f s ⁄ 2 ,
is filtered out and, ideally, the SNR within this frequency range is now given by [37]
SNR ≈ 6.02N + 1.76 + 10 log OSR (1-2)
81
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
The OSDAC can obviously be designed for low noise and high resolution if the over-
sampling ratio is high due to the term: 10 log OSR in (1-2). However, the increase of SNR
is still only 3dB for every doubling of the OSR and when increasing the dynamic range a
higher resolution of the DAC still is required to meet the specifications on the linearity.
In a 14-bit DAC with an OSR = 4, the linearity of the converter must correspond to a 15-
bit DAC [37].
L DAC LP
Figure 1-1: Naive use of oversampling. The input signal is oversampled L (= OSR) times. Images that
occur due to the oversampling is filtered out using the lowpass filter.
By using noise shaping with a modulator (Fig. 1-2), the linearity and the SNR of the con-
verter may be improved significantly, as is further described in Chapter 2. In fact, the
improvement in SNR is approximately given by (in dB)
( 20 ⋅ M + 10 ) log OSR (1-3)
L Σ∆ DAC LP
Figure 1-2: Oversampling DAC using modulation to improve linearity and signal-to-noise ratio.
The modulator is purely digital and it reduces the number of bits of the signal. This intro-
duces a higher quantization noise which is highpass filtered and the original signal infor-
mation is lowpass (or allpass) filtered. The reduced number of bits also relaxes the
requirements on the DAC used in the converter and hence the number of analog building
blocks may be decreased. By reducing the number of bits to only one single bit, i.e., pulse
code modulation (PCM), the DAC can be designed to be completely linear since it has
only two threshold values [11]. Noise shaping is also referred to as sigma-delta (Σ∆) or
delta-sigma (∆Σ) modulation.
The ADSL standard as described in Part A: Overview: Digital-to-Analog Converter
Design has a bandwidth of approximately 1.1MHz [83]. Since the OSDAC contains a few
number of analog elements, the oversampling ratio can be large. With an oversampling
ratio of 32, the sampling frequency is approximately 71MHz for the OSDAC for ADSL
(Chapter 3) and basically there are no bigger problems designing digital circuitry for these
speeds. The OSDAC is implemented in Ericsson’s 0.6µm CMOS process.
In this part we focus on and describe a behavioral-level design of an OSDAC where the
trade-offs between different building blocks are highlighted and discussed.
82
2. Oversampling DAC Structures
2 Oversampling DAC
Structures
Several different oversampling converter structures can be found [11, 37, 41, 46, 56, 59,
60, 63]. We show a general topology (Fig. 2-1) which is divided into a number of basic
building blocks; interpolation filter, modulator, M-bit DAC, and analog filter.
N bits @ f OSR
X2(kTo)
Interpol. Modu- DAC Analog
X1(kTs) Filter lator (M bits) X4(kTo) LP filter Aout(t)
X3(kTo)
N bits @ f s M bits @ f OSR
Figure 2-1: General structure of the oversampling digital-to-analog converter. Interpolation filter,
Modulator, DAC, analog lowpass filter (smoothing filter). The interpolation filter fulfils
the oversampling of the signal and the modulator the reduction of bits in the circuit.
N bits @ f OSR
X (kT )
Interpol. 2 o Modu- Semi-digital Analog
X1(kTs) Filter lator X3(kTo) FIR filter X4(kTo) LP filter Aout(t)
N bits @ f clk 1 bit @ f OSR
Figure 2-2: Structure of the oversampling digital-to-analog converter when reducing the number of
bits to one.
When letting the modulator reduce the number of bits to only one single bit, see Fig. 2-
2, the DAC can be replaced by a semi-digital FIR filter (Sec. 2.3), that both converts from
83
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
a digital to an analog representation and filters the signal as well [34, 63]. The semi-digital
FIR filter can be considered to be a number of cascaded single-bit DACs with different
reference levels. With this method the analog smoothing filter can be designed with less
complexity. However, the modulator must be designed more carefully with respect to sta-
bility since the loop gain is high.
a)
fs / 2 fs
b) Interpolation
filtering
c)
Shaped noise
d)
Lowpass filtering
e)
fs / 2
Figure 2-3: Frequency spectrum for different stages of the oversampling with noise shaping. In this
example OSR=4. a) Original spectrum. b) Oversampling spectrum. c) Noise is shaped. d)
Lowpass filtering the original signal e) Wanted output spectrum.
84
2. Oversampling DAC Structures
OSR = L 1 ⋅ L 2 ⋅ … ⋅ L N (2-1)
y ( kT ) = x ( kT ) k = m⋅L (2-2)
0 k≠m⋅L
where y ( kT ) is the output signal, x ( kT ) is the input signal, L is the oversampling ratio,
m = 0, 1, 2, … is an integer, and T is the update time interval. In the frequency domain
(2-2) corresponds to
The total transfer function, H OSR ( z ) , of the filters in Fig. 2-4 is given by
where H i ( z ) are the subfilters of the multi-stage filter. By using interpolation in stages or
frequency masking techniques [36] the filters H i ( z ) may be designed easier and with less
complexity and less power consumption. A simple structure of a one-stage interpolation
filter is shown in Fig. 2-5. It is an FIR filter with K taps and the output is given by
The implementation of long filter coefficients may cost much and therefore filter struc-
tures with simple coefficients are wanted.
x(nT)
T/OSR T/OSR T/OSR
a0 a1 a2 aK-2 aK-1
y(nT)
If the oversampling ratio is high, the interpolation filter has to be narrow banded which
increases the number of coefficients in the filters and the complexity [36].
85
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
2.2 Modulator
Basically, the modulator in an OSDAC is designed to reduce the number of bits repre-
senting the signal, i.e., to use a pulse code modulation (PCM). This implies that the quan-
tization noise will increase, but the noise can be shaped in such a way that it is highpass
filtered. The transfer function from the noise to the output will be referred to as the noise
transfer function (NTF) and for the signal we will refer to the signal transfer function
(STF). The STF is either an allpass or lowpass filtering function.
Generally, the modulator contains a feedback loop and IIR filters to shape the noise. The
feedback is of two kinds:
• Signal feedback (Sec. 2.2.1)
• Error feedback (Sec. 2.2.2)
The filters used in the modulator are different in the two structures. Dependent on the
number of bits at the output the modulator structures are divided into two groups:
• Multi-bit modulators
• Single-bit modulators
In this thesis, we focus on the single-bit modulator, which has a large loop gain and
therefore the stability must be carefully considered [56, 60]. In the single-bit modulator,
a certain value at the input is represented by the average value of the pulses at the output,
hence the average duty cycle. Hence the maximum value of the input is represented by a
stable maximum output value on the output. The mean value of the input is represented
by a square waved signal with 50% duty cycle. In Fig. 2-6 the simulation result for a first
order modulator using signal feedback (Sec. 2.2.1) is shown. The input is sweeping from
a value near the minimum amplitude to a value near the maximum (dashed line) and the
output signal is described by the square waves (solid line).
0.8
0.6
0.4
Normalized Amplitude
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 50 100 150
Sequence Index
Figure 2-6: The average value of the output approximates the input signal. The simulation results are
achieved with a first order signal feedback modulator.
86
2. Oversampling DAC Structures
G(z)
H (z)
STF ( z ) = ------------------------------------- (2-10)
1 + H (z) ⋅ G(z)
1
NTF ( z ) = ------------------------------------- (2-11)
1 + H (z) ⋅ G(z)
The signal should be lowpass (or allpass) filtered and the noise highpass filtered. This
states the requirements on the filters H ( z ) and G ( z ) , and the order of the modulator is
determined by the filter orders.
Assume that we want the STF to be described by a delay only, hence an allpass filter
STF ( z ) = z –1 (2-12)
and for the first order modulator the NTF is described by the highpass filter
NTF ( z ) = ( 1 – z –1 ) (2-13)
87
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
z –1
G ( z ) = 1 and H ( z ) = ---------------
- (2-14)
1 – z –1
For the second order modulator the NTF is
NTF ( z ) = ( 1 – z –1 ) 2 (2-15)
Using (2-12) and (2-15) in (2-10) and (2-11), the filters are set to
z –1
G ( z ) = 2 – z –1 and H ( z ) = ----------------------
- (2-16)
( 1 – z –1 ) 2
In Fig. 2-8 the NTFs for the first and second order modulator are shown. The second
order modulator (dotted line) shows a better noise attenuation at lower frequencies than
the first order modulator (solid line).
−5
|NTF(z)|2 [dB]
−10
−15
−20
−25
−30
−35
0 0.1 0.2 0.3 0.4 0.5
Normalized Frequency
Figure 2-8: NTF of a first order (solid line) and second order (dotted line) modulator.
The NTF for the second order modulator (when neglecting the sinc weighting from sam-
ple-and-hold elements) is
f 2
– j ⋅ 2π -----------
NTF ( f ) = ( 1 – z –1 ) 2 j ⋅ 2π ⋅ ( f ⁄ f OSR ) = 1 – e f OSR
=
z=e
f
– j ⋅ 2π -----------
⋅ sin2 π ------------
f
= –4 ⋅ e f OSR
(2-17)
f OSR
88
2. Oversampling DAC Structures
where it is assumed that the OSR is high. Within the interesting frequency range, f N , the
noise power can be calculated by using the NTF and the known, original, quantization
noise error PSD
2
A ref
1
S q ( f ) = ------------ ⋅ ---------- (2-19)
f OSR 12
where A ref is the reference step value. The total power of the shaped noise within the
Nyquist band is found by combining (2-18) and (2-19)
fN fN
where
f OSR = OSR ⋅ 2 f N (2-21)
1--- ⋅ 2 2N – 2 ⋅ A 2
P 2 ref
SNR = 10 log ---------------- = 10 log ---------------------------------------
signal
- =
P noise A 2 ⋅ ----- π4 1
ref 60- ⋅ ------------- OSR 5
-
From (2-23) it is seen that for each doubling of the OSR, the SNR is improved by 15dB,
corresponding to 2.5 bits. The improvement is approximately given by M + 0.5 bits for
each doubling of the OSR [56]. M is the order of the modulator. For the single-bit mod-
ulator the linearity is also guaranteed since a single-bit DAC is used at the output [37].
89
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
-
+
G(z) q(nT)
Figure 2-9: Basic principle of an error feedback modulator.
STF ( z ) = H ( z ) (2-27)
NTF ( z ) = 1 – H ( z ) ⋅ G ( z ) (2-28)
Since the signal should be lowpass or highpass filtered and the noise should be highpass
filtered, the requirements on the filters H ( z ) and G ( z ) are set. Assume that the STF is
described by a unity function
STF ( z ) = H ( z ) = 1 (2-29)
For the first order modulator the NTF is described by
NTF ( z ) = ( 1 – z –1 ) (2-30)
and by using (2-29) in (2-28) with (2-30) we have
G ( z ) = z –1 (2-31)
For the second order modulator the NTF is
NTF ( z ) = ( 1 – z –1 ) 2 (2-32)
and therefore
G ( z ) = z –1 ( 2 – z –1 ) (2-33)
The feedback filters have practically the same structure as those for the signal feedback
configuration and the gain in SNR as function of the OSR is also the same.
90
2. Oversampling DAC Structures
z –1
A ( z ) = ---------------
- (2-34)
1 – z –1
aN/2 a1
x(nT) + - y(nT)
A(z) A(z) A(z)
- - -
bN bN-1 b1
Figure 2-10: General structure of a higher order modulator using signal feedback structure.
where A = A ( z ) is the filter given by (2-34). The zeros of the NTF in (2-36) are
z 1, 2 = 1 ± j a 1 and z 3, 4 = 1 ± j a 2 (2-37)
and the a i :s are used to move the poles of the NTF to improve the stability. The poles are
given by complex expressions and the overall complexity of the synthesis of the filter
structures increases dramatically with increased modulator order. Simulation results of
stability issues are also discussed for the implemented fifth order modulator in Sec. 3.1.2.
91
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
is given by
y ( nT ) = a 0 ⋅ w ( nT ) + a 1 ⋅ w ( nT – T ) + a K – 1 ⋅ w ( nT – ( K – 1 )T ) (2-38)
where K is the number of taps in the FIR filter and T = 1 ⁄ f s is the clock cycle time.
The coefficients are generated with analog components, for example current sources in a
current-steering version [45]. The output, y ( nT ) , is a pulse-shaped signal since in general
the multiplication by the coefficients is done with sample-and-hold elements. This will
generate images that have to be filtered out by the analog smoothing filter.
w(nT)
T T T
a0 a1 a2 aK-2 aK-1
y(nT)
The number of coefficients in an FIR filter is inversely dependent on the width of the
transition band [36], as
1 1
K ∼ ------------ = --------------------------- (2-39)
∆ωT ωa T – ωc T
where ω c T is the normalized cut-off frequency of the passband and ω a T is the normal-
ized stopband frequency. If the coefficients are signal-independent, naturally no distortion
will be introduced, only the filter’s transfer function will change if there are matching er-
rors in the current sources. With an FIR filter a linear phase is guaranteed as well. Long
FIR filters naturally introduce a large delay through the shift register (Fig. 2-11).
FIR CT LP
Figure 2-12: Images and noise are attenuated using the continuous-time filter
92
2. Oversampling DAC Structures
FIR DT LP CT LP
Figure 2-13: Images and noise are attenuated further by using a discrete-time filter (DT LP) between
the FIR filter and the continuous-time filter (CT LP).
Naturally, the structure of the filters (Butterworth, Cauer, etc.) and circuit-technique
(current-mode, voltage-mode, etc.) are dependent on the application and the specification.
Sinc weighting
FIR Filter
Analog LP
fN fOSR/2 fOSR
Figure 2-14: Filtering stages occurring in the OSDAC, illustrated in the frequency domain.
As is sketched in Fig. 2-14 there are actually three filtering functions occurring at the
back-end of the OSDAC; the semi-digital FIR filter, the sinc weighting of the output spec-
trum and the analog continuous-time lowpass filter. When determining the order of the
analog filter, these aspects should be considered. If the OSR is high, the influence of the
sinc weighting in the signal band is very low.
To find the suitable trade-off the different building blocks are modeled and simulated on
a higher-level and at least a local optimum solution may be found with the models used
[1]. The circuit implementation will not give the same result, but it will be close to the
local optimum found by behavioral-level modeling and simulation.
All the design issues can briefly be summarized for the different building blocks as
• Using higher order interpolation filter introduces
→ higher complexity, filter structures, filter orders, etc.
→ higher power consumption
→ decreased out-of-band noise (and attenuation of images)
→ relaxed requirements on the modulator
93
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
94
3. A CMOS Oversampling DAC for ADSL Applications
3 A CMOS Oversampling
DAC for ADSL Applications
In this chapter the design of a 14-bit 3V-5V CMOS OSDAC for ADSL applications is
presented. The work is done by J. J. Wikner (Linköping University), Y. Gao (ESDLab,
KTH), and N. Tan (GlobeSpan Semiconductor). The converter is fabricated in Ericsson’s
0.6µm CMOS process.
As a very important design issue, the simulation of the whole converter on higher level
will be discussed. We show the impact of different filter orders (interpolation, FIR and
analog), and how this applies to the ADSL standard.
In the following sections the building blocks of the OSDAC (Sec. 3.1), simulation results
(Sec. 3.2), and the chip implementation (Sec. 3.3) are presented. Finally, in Sec. 3.4, the
drawbacks and the possible improvements of the implementation are discussed.
z–1 J
H 1 ( z ) = ----------- = ( 1 – z –1 ) J (3-1)
z
1 J z –1 J
H 2 ( z ) = ----------- = ---------------
- (3-2)
z – 1 1 – z –1
X(z) Y(z)
L z-1 z-1
95
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
The exponent J from (3-1) and (3-2) is set to J = 0, 1, 2 in the different versions. For
even higher J :s the stability cannot be guaranteed in the simulations due to the pole on
the unit circle at dc level. The total transfer function of the interpolation filter, H ip ( z ) ,
describes a lowpass filtering and is given by
z –1 J
H ip ( z ) = H 1 ( z L ) ⋅ H 2 ( z ) = ( 1 – z – L ) J ⋅ ---------------
- (3-3)
1 – z –1
where L is the oversampling ratio and L = OSR . Using a series expansion, (3-3) can be
rewritten as
z –1 J
H ip ( z ) = ( 1 + z –1 + … + z –( L – 1 ) ) J ⋅ ( 1 – z –1 ) J ⋅ ---------------
- =
1 – z –1
= ( z –1 + z –2 + … + z –L ) J (3-4)
and the poles in the expression are cancelled. The magnitude response of the filter func-
tion (3-4) with J = 1 is displayed in Fig. 3-2. The attenuation of the signal at f N (dashed
line) is approximately 3.9dB.
30
20
Magnitude [dB]
10
−10
−20
−30
0 0.1 0.2 0.3 0.4 0.5
Normalized Frequency
Figure 3-2: Magnitude response of the interpolation filter as in (3-4) with J=1. The original Nyquist
bandwidth is indicated with the dashed line.
In the final version of the OSDAC chip, no interpolation filter was implemented.
3.1.2 Modulator
In the OSDAC single-bit, signal feedback modulators of the fourth and fifth order were
investigated. They are designed by Y. Gao (KTH), and a closer description of the design
of the fifth order modulator is presented in this section. The modulator’s structure is
shown in Fig. 3-3, it uses a 14-bit two’s complementary input at the frequency
f OSR = 1.104 ⋅ 64 ≈ 71 MHz.
96
3. A CMOS Oversampling DAC for ADSL Applications
b 1 = 2 11 , b 2 = 2 10 , b 3 = 2 8 , b 4 = 2 5 , b 5 = 2 2 , a 1 = 2 –7 , a 2 = 2 –8
a2 a1
x(nT) - - y(nT)
+ - A(z) -
A(z)
-
A(z)
-
A(z)
-
A(z)
b5 b4 b3 b2 b1
The filters, A ( z ) , in Fig. 3-3, are discrete-time integrators with the transfer function
z –1 1
A ( z ) = ---------------
–
- = -----------
1
(3-5)
1–z z–1
where A = A ( z ) from (3-5). From (2-37) the zeros of the NTF are
z 1, 2 = 1 ± j a 1 and z 3, 4 = 1 ± j a 2 (3-7)
Without a2 and a3
80
60
40
20
−20
−40
−60
−80 0 1 2 3 4 5
10 10 10 10 10 10
Figure 3-4: Output amplitude spectrum of the modulator’s output without the complex NTF zeros.
The input signal frequency is 431.25kHz and the sampling frequency is 70.66MHz.
The amplitude spectrum of the modulator’s output when applying a sinusoidal with sig-
nal frequency 431.25kHz is shown in Fig. 3-4. The SNDR is approximately 63dB.
97
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
With the a i parameters the SNDR is improved. The zeros further attenuate the noise
within the signal band and a sharper transition band is achieved. The simulation result
when applying a sinusoidal at a frequency of 431.25kHz is shown in Fig. 3-5 where the
SNDR is found to be approximately 79dB.
With a2 and a3
80
60
40
20
−20
−40
−60
−80 0 1 2 3 4 5
10 10 10 10 10 10
Figure 3-5: Output amplitude spectrum with the two complex pairs of NTF zeros. The input signal
frequency is 431.25kHz and the sampling frequency is 70.66MHz
The stability of the modulator was simulated using a root locus shown in Fig. 3-6. With
the locus the limits of the feedback values, b i , are found and especially the gain of the
feedback loop can be determined.
1.5
0.5
−0.5
−1
−1.5
−1.5 −1 −0.5 0 0.5 1 1.5
Figure 3-6: Root locus of the modulator when varying the feedback parameters.
The layout of the modulator is synthesized using COMPASS tools and is shown in Fig.
98
3. A CMOS Oversampling DAC for ADSL Applications
Figure 3-7: Layout view of the fifth order modulator generated with COMPASS.
w(nT)
D D D
y(nT)
Figure 3-8: Current-steering implementation of a semi-digital filter with coefficient length K.
The digital delay line or the shift register is constructed using dynamic CMOS gates [81]
with help from Magnus Karlsson (Linköping University). The D-latch (Fig. 3-9c) is
formed by using a P-gate and an N-gate as shown in Fig. 3-9a-b). The D-latch uses only
99
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
one single clock phase which simplifies the clock distribution. The outputs are comple-
mentary and hence suitable for controlling the differential current switches.
a) b) c)
Figure 3-9: Schematic of a) P-gate and b) N-gate, and c) D-latch formed by a P-gate and N-gate.
The latches are sensitive to the capacitive load on the outputs. A constant load is guar-
anteed by letting all current switches be equally large. The layout of the D-latch can be
made rather compact. In the design of the FIR filter the latches have been laid out (Fig. 3-
10) so that they are as wide as the current switches and the unit current source, hence a
regular layout is possible.
clk
D+ D- 2)
Figure 3-10: Layout view of a D-latch including the 1) P-gate and 2) N-gate.
The coefficients as shown in Fig. 3-8 are created by using a number of unit current
sources in parallel according to the same principle as in the design of the converters in
Part B: Nyquist-Rate Digital-to-Analog Converters for VDSL Applications and the
matching issues apply in the same way. In this implementation single cascode PMOS
transistors are used. The layout of a unit current source is shown in Fig. 3-11. M1 and M2
are given by the sizes ( W ⁄ L ) 1 = 2 ⁄ 8 and ( W ⁄ L ) 2 = 2 ⁄ 2 [µm].
VDD M1 M2 Output
Gates
Figure 3-11: Layout view of a current source constructed by two PMOS transistors. M2 is the cascode.
100
3. A CMOS Oversampling DAC for ADSL Applications
For the FIR filter, the cut-off frequency is approximately f c = 1.15 ⋅ f N ≈ 1.27 MHz
and the attenuation in the stopband is approximately 17dB. The FIR filter coefficients
have been generated with the Remez algorithm in MatLab. Since the original coefficients,
a i , have been amplified and truncated to integer values, ã i , there is no longer an equi-
ripple in the stopband. Proper values are found as
max a i
ã i = J ⋅ ------------------- + 0.5 ⋅ sgn { a i } (3-8)
min a i
where J is an amplification value which determines the size of the truncation error. In Fig.
3-12 the filter magnitude response for the original filter (219 taps) and the truncated ver-
sion is shown. The FIR filter in the implemented version of the OSDAC uses
• 2390 effective unit current sources in 219 nonzero taps
• 152 unit current sources in the largest tap
A large number of unit current sources in the taps increases the accuracy, since J is
larger, but it also generates a higher current which makes the design of current switches,
wires, supply, etc., harder.
40
Increased
Power [dB]
attenuation
20 in stopband
−20
−40
0 0.5 1 1.5 2 2.5 3 3.5 4
Frequency 7
x 10
Figure 3-12: 219 tap FIR filter magnitude response. Characteristics with and without truncated filter
coefficients are shown.
The switches are differential and therefore also negative filter coefficients can be real-
ized. For a current source corresponding to a negative coefficient the outputs of the dif-
ferential switch is simply cross-connected to the OSDAC outputs, as shown in Fig. 3-13.
All current switches in the OSDAC are equally large with eight NMOS transistors in par-
allel for each channel. This also ensures equal load for all D-latches. The layout of one of
the switches is shown in Fig. 3-14. NMOS transistors are used since they have higher
transconductance parameter (higher mobility).
101
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
φ0 φ0 φ1 φ1 φM φM
I+
I-
a0: Positive a1: Negative aM: Negative
Coefficient Coefficient Coefficient
Figure 3-13: Differential current switches in the FIR filter. Negative coefficients are realized by cross
connecting the outputs.
G ON ≈ 8β ⋅ ( V φ – V T – V C ) (3-9)
where β is the transconductance parameter for one single NMOS transistor. V φ is the
gate or switch voltage on the transistors, V C is the output node of the current sources and
V T is the threshold voltage of the transistors. Since the output node is varying with the
signal, the source-bulk voltage, V SB , varies as well. This forces the threshold voltage to
be slightly non-linear.
φ I+
Ii
Shielding I-
φ
Ground
Figure 3-14: Layout view of differential current switch used in FIR filter.
The average output current, I out , of the FIR filter is given when the input is alternating
between “0” and “1” with a 50% duty cycle, as
hence the average power for one channel (standby for a zero input in two’s complement)
over a 50Ω load is given by
6
P out = I out 2 ⋅ 50 = 86.86 ×10 ⋅ I unit
2 (3-11)
102
3. A CMOS Oversampling DAC for ADSL Applications
The output current of the unit current source is dependent on the reference current. For
the case with I unit ≈ 1.65µA we have from (3-10) and (3-11)
The impulse response of the semi-digital FIR filter from a circuit-level simulation is
shown in Fig. 3-15. The output current of the unit current source is approximately
I unit ≈ 1.65µA and the output is doubly terminated with 50Ω. Since the start-up phase of
the simulation was too long, the figure only displays 172 of 219 values.
4.4
4.35
Differential output current [mA]
4.3
4.25
4.2
4.15
4.1
4.05
3.95
0 200 400 600 800 1000 1200
Time [us]
Figure 3-15: Impulse response from a circuit-level simulation of the semi-digital FIR filter.
Since the currents originally were to be terminated with an operational amplifier (input
stage to the analog smoothing filter) with low input impedance, the total output imped-
ance of the FIR filter has not been investigated. The analysis is however similar to the case
with the binary weighted sources in the Nyquist-rate converters.
103
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
I-in R
V+out
I+out
I-out
I+in V-out
R
C
Figure 3-16: First order differential lowpass filter.
From the ADSL specification the cut-off frequency is given and therefore
1 –9
RC = -----------------------------------6- ≈ 144 ×10 (3-15)
2π ⋅ 1.104 ×10
We also know that the dc voltage drop over the resistance, R , can not be too large for
low voltage applications. With an input dc current, I dc , of 2.2mA and a maximum voltage
drop of V drop, max = 1 V, we have that
–3
I dc ⋅ R < V drop, max or R < 1 ⁄ 2.2 ×10 ≈ 455Ω (3-16)
which is far too large to implement on chip since the area would be approximately 1mm2
with today’s CMOS technology. Another filter structure, where the circuit is not used as
an I/V converter, could be used as on-chip filter (see Sec. 3.4). However, no on-chip an-
alog filter was implemented in the fabricated OSDAC chip.
104
3. A CMOS Oversampling DAC for ADSL Applications
of the filters are slightly above 1.104 MHz. The ripple in the passband is mainly deter-
mined by the FIR filter and the round-off errors of the coefficient truncation. The interpo-
lation filters are sinc filters as shown in Sec. 3.1.1. The analog filters are Butterworth
filters as in Sec. 3.1.4. In Fig. 3-17 we see that the images from 1.104MHz to 2.208MHz
are very large and that the noise is not attenuated as much as wanted. The ripple in the
passband of the 210 carriers are around the specified level (-36.5dB).
ADSL Specifications
−20 Images
within
New ADSL fN and 2*fN.
Specification −40
−60
Noise and
distortion
components.
PSD
−80
−100
210 carriers
−120
−140 3 4 5 6 7
10 10 10 10 10
Frequency
Figure 3-17: Output spectrum without interpolation and analog filters.
By using a second order interpolation filter the images are slightly attenuated as shown
in Fig. 3-18. The noise is not affected since it origins in the modulator. The attenuation of
the noise by the FIR filter is not large enough. It is also seen that the interpolation filters
cause a quite large attenuation near the passband edge. The FIR could be designed to can-
cel this effect by having a higher gain near the passband edge.
ADSL Specifications
−20
Drop, due to
interpolation
−40 filter.
−60
PSD
−80
−100
−120
−140 3 4 5 6 7
10 10 10 10 10
Frequency
Figure 3-18: Output spectrum with a second order interpolation filter and without analog filter.
105
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
By using the first order analog smoothing filter (Fig. 3-19) the noise is attenuated and
the images are slightly attenuated. We see an even larger attenuation of the high frequency
carriers within the passband, due to the too low cut-off frequency in the analog filter.
ADSL Specifications
−20
−40
−60
PSD
−80
−100
−120
−140 3 4 5 6 7
10 10 10 10 10
Frequency
Figure 3-19: Output spectrum with second order interpolation filter and a first order analog filter.
With a third order analog filter (Fig. 3-20) the specification is still not met, the noise is
almost attenuated to the specified level (-90dB).
ADSL Specifications
−20
−40
−60
PSD
−80
−100
−120
−140 3 4 5 6 7
10 10 10 10 10
Frequency
Figure 3-20: Output spectrum with a second order interpolation filter and third order analog filter.
By increasing the FIR filter length to 324 taps the attenuation in the stopband is
increased, hence the noise is attenuated as seen in Fig. 3-21. The specification is however
still not met. There are hard requirements and the design of the filters has to be carefully
done.
106
3. A CMOS Oversampling DAC for ADSL Applications
ADSL Specifications
−20
−40
−60
PSD
−80
−100
−120
−140 3 4 5 6 7
10 10 10 10 10
Frequency
Figure 3-21: Output spectrum with a 324 tap FIR filter, second order interpolation filter, and a third
order analog filter.
It should be noted that the simulation results presented in the figures above are compared
with a newer ADSL standard and the fabricated OSDAC has been designed to meet the
specification of an older ADSL standard. As was discussed in Part A: Overview: Digital-
to-Analog Converter Design, the newer ADSL standard sets much tougher requirements
on the attenuation of higher frequency components and especially the frequency range
from 1.104MHz to 2.208MHz.
It is easy to understand the complexity of the choices that have to be taken when choos-
ing the filter and modulator orders.
107
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
FIR Filter
Analog I/O
Digital I/O
Modulator
Decoupling
Capacitors
Figure 3-22: Layout view of the whole OSDAC with modulator and FIR filter included.
108
3. A CMOS Oversampling DAC for ADSL Applications
Assume that the original FIR filter, H 0 ( z ) , has the following specification (2-39)
With an attenuation in the stopband of approximately 17dB the number of taps is 219
(Sec. 3.1.3). Using an IFIR filter, H 1 ( z 2 ) , with ω c T ≈ 2.30 f N and ω s T ≈ 2.94 f N . With
the same attenuation in the stopband, the number of taps is around 109 (from Remez’
algorithm in MatLab). In Fig. 3-23 this is illustrated with a filter interpolated at an order
of eight, H 2 ( z 8 ) . The behavior within the passband is the same for both filters. The pass-
band of the IFIR filter is repeated over the frequency range and the shaped noise from the
modulator will not be filtered out. Therefore, the attenuation of the analog filter at the out-
put must be higher to remove these components and the transition band must be sharper
as well. Not that the order of interpolation may not be too large, since the design of the
following filters will be too hard.
Magnitude Response of FIR and IFIR filter
10
−10
Magnitude [dB]
−20
−30
−40
−50
−60
0 5 10 15 20 25 30 35
Frequency [MHz]
Figure 3-23: IFIR filter (interpolation order = 8) vs. FIR filter. The behavior within the passband is the
same for both filters. The IFIR filter has an offset of 5dB to illustrate the function.
It can be seen from simulation results that the order of the analog filter (to still meet the
ADSL specifications) is approximately proportional to the interpolation ratio in the IFIR
filter. In Table 3-1 the approximate filter orders are summarized. As analog filter, a But-
terworth filter is used. It is seen that the specifications on the analog filter may become
much tougher. It should also be noted that the delay through the IFIR filter still is the same
as for the original FIR filter.
The noise transfer function of the modulator could also be designed to have less noise
within the passband images that arise due to the interpolation [41].
109
Part C: Oversampling Digital-to-Analog Converters for ADSL Applications
To further simplify the design of the analog filter, the dc currents from the outputs of the
FIR filter can be removed by using current sinks implemented with NMOS transistors,
Fig. 3-24. The ac currents are fed into the analog LP filter. As was discussed in Sec. 3.1.4
the dc voltage drop over the resistance (Fig. 3-16) now can be removed and the resistance
can be chosen to be larger, hence the capacitance can be made smaller (3-15) and a chip
implementation may be possible.
Semi-digital FIR
Ipos Ineg Ipos,ac
LP
Ineg,ac
IDC IDC
Figure 3-24: Removing the dc current from the FIR filter output relaxes the complexity of the design
of the analog LP filter.
As was briefly mentioned an additional switched-current (SI) filter could also be used to
filter the signal [45, 71]. A frequency masking technique using an FIR filter and SI filter
(which is discrete-time) would relax the specifications on the analog filter as well.
In the trade-off discussion it was also mentioned that the complexity and power con-
sumption will increase with an interpolation filter and on-chip analog filter. This must
also be considered when changing the structure of the OSDAC.
110
Conclusions
111
Conclusions
Measurement results from the oversampling converter have not been presented and
therefore some conclusions of the implementation are missing. The same issue with
matching of current sources can be applied to this converter as well, since the FIR filter
was implemented with unit current sources. Here, we have discussed some techniques to
further improve the performance to meet the specifications of the newer ADSL standards.
We have discussed the important issues of simulating and measuring circuits. It is found
that small errors in test vectors may generate large errors in the output signal of the con-
verter. From this we conclude that it is important to have good input data to trust the out-
put signal as well as be able to make the correct conclusions from this.
In general, we have pointed out the behavioral-level simulations as a very useful and
important stage of the design. It is important that the designer is aware of all stages of the
development of the chip. By using good and simple models, it is possible to find funda-
mental limits on the performance and therefore also be able to focus on improving the
components that are bottlenecks in the systems. These higher-level simulation has espe-
cially been used during the design of the oversampling converter, where extensive Mat-
Lab models have been developed.
112
References
1
References
2
References
[26] S. Haykin, Digital Communication, John Wiley & Sons, New York, NY, USA, 1988,
ISBN 0-471-633775-0
[27] B. Hedberg, “Wideband Radio System,” Private communication, Ericsson Radio
Systems AB, Sweden
[28] P. Hendriks, “Specifying Communication DACs,” IEEE Spectrum, vol. 34, no. 7,
pp. 58-69, July 1997
[29] P. Hendriks, “Tips for Using High-Speed DACs in Communications Design“, IEEE
Electronic Design, no. 2, pp. 112-8, Jan. 1998
[30] B. G. Henriques and J. E. Franca, “High-Speed D/A Conversion with Linear Phase
sinx/x Compensation,” in Proc. of the 1993 IEEE Int’l Symp. on Circuits and Systems
(ISCAS’93), vol. 2, pp. 1204-7, Chicago, IL, USA, May 3-6, 1993
[31] B. G. Henriques and J. E. Franca, “A High-Speed Programmable CMOS Interface
System Combining D/A Conversion and FIR Filtering,” IEEE J. of Solid-State
Circuits, vol. 29, pp. 972-7, no. 8, Aug. 1994
[32] B. G. Henriques, K. Kananen, J. E. Franca, and J. Rapeli, “A 10 Bit Low-Power
CMOS D/A Converter with On-Chip Gain Error Compensation,” in Proc. of the 1995
IEEE Custom Integrated Circuits Conf. (CICC’95), pp. 215-8, Santa Clara, CA, USA,
May 1995
[33] M. Ismail and T. Fiez, Analog VLSI: Signal and Information Processing, McGraw-
Hill, New York, NY, USA, 1994, ISBN 0-07-113387-9
[34] C. Jansson, CMOS Signal Conditioned Sensor Arrays and A/D-D/A Conversion
Techniques, Linköping Studies in Science and Technology, Dissertation no. 377,
1995, ISBN 91-7871-521-0
[35] H. T. Jensen and I. Galton, “A Low-Complexity Dynamic Element Matching DAC for
Direct Digital Synthesis,” IEEE Trans. on Circuits and Systems II: Analog and Digital
Signal Processing, vol. 45, no. 1, pp. 13-27, Jan. 1998
[36] H. Johansson, Synthesis and Realization of High-Speed Recursive Digital Filters,
Linköping Studies in Science and Technology, Dissertation no. 534, 1998, ISBN 91-
7219-207-0
[37] D. A. Johns and K. Martin, Analog Integrated Circuit Design, John Wiley & Sons,
New York, NY, USA, 1997, ISBN 0-471-14448-7
[38] B. E. Jonsson, “Design of Power Supply Lines in High-Performance SI and Current-
Mode Circuits,” in Proc. of the 15th Norchip Sem., pp. 245-50, Tallinn, Estonia, Nov.
10-11, 1997
[39] I. H. H. Jørgensen and S. A. Tunheim, “A 10-bit 100MSamples/s BiCMOS D/A
Converter,” Analog Integrated Circuits and Signal Processing, vol. 12, pp. 15-28,
1997
[40] P. Ju, K. Suyama, P. Ferguson Jr., and W. Lee, “A Highly Linear Switched-Capacitor
DAC for Multi-Bit Sigma-Delta D/A Applications,” in Proc. of the 1995 IEEE Int’l
Symp. on Circuits and Systems (ISCAS’95), vol. 1, pp. 9-12, Seattle, WA, USA, April
29 - May 3, 1995
[41] J.-W. Kim, B.-M. Min, J.-S. Yoo, and S.-W. Kim, “An Area-Efficient Sigma-Delta
DAC with a Current-Mode Semidigital IFIR Reconstruction Filter,” in Proc. of the
1998 IEEE Int’l Symp. on Circuits and Systems (ISCAS’98), Monterey, CA, USA, May
31 - June 3, 1998
3
References
4
References
5
References
6
Licentiate Theses
Division of Electronics Systems
Department of Electrical Engineering
Linköpings universitet
Sweden
Melander J.: Design of SIC FFT Architectures, Linköping Studies in Science and
Technology, Thesis No. 618, 1997.
Fig. 2. Current source cell topologies. (a) Basic unary current source. (b)
Cascoded unary current source.
II. STATIC PERFORMANCE
Static and dynamic performances of current-steering DACs
the current source. This impedance boosting is needed in order
are mostly determined by the current sources accuracy, nonin-
to fulfil requirements for high-resolution high-bandwidth cur-
finite output impedance, and switching time. In the following,
rent-steering DACs [2].
the dependencies of the DAC static performance on the cur-
DAC performance is specified both through static parameters, rent sources’ accuracy and output impedance constrained by the
namely integral nonlinearity (INL), differential nonlinearity transistors’ operating region are presented. Afterwards, the im-
(DNL), and parametric yield, as well as dynamic parameters, plications of those static and statistical constraints in the sizing
namely glitch energy, settling time, and spurious-free dynamic design procedure of the current cell circuit are presented and
range (SFDR) [3]. Static performance is mainly dominated compared with other previously published approaches.
by systematic and random mismatch errors. Systematic errors
caused by process, temperature, and electrical slow variation
gradients are almost cancelled by proper layout techniques [4]. A. Current Source Accuracy
Random errors are determined solely by random mismatch due The transistor CS of Fig. 2 sets the cell’s current. Based on the
to fast process variation gradients. statistical mismatch model of [12], a relation exists between the
The design of a DAC is carried out at three levels: architec- CS transistor area, its overdrive voltage, and its relative accuracy
ture design, circuit design, and physical design. In the usual de- (we will refer in the following to the LSB current source)
sign procedure, the degree of segmentation ( over ) is de-
cided at the architecture level by combining static and dynamic
specifications and overall minimum area requirements [5]–[7].
(1)
Subsequently, the basic current cell circuit (usually the LSB cur-
rent source) is designed by determining the sizes and bias volt-
ages for the different transistors and the voltage levels required where is the cell’s current, and are the CS
for the switches’ control signals. Most of the DAC static and transistor width and length, respectively, and are
dynamic performance is determined by the performance of the mismatch process constants for the transistor large signal gain
basic current source cell. The circuit design of the rest of the and threshold voltage, respectively, is the CS overdrive
components within the DAC architecture is mainly influenced voltage, i.e., , where is the CS gate-to-source
by speed requirements. The digital binary-to-thermometer de- voltage and is the threshold voltage, and is the
coder must be fast enough so that the converter speed is solely variance of the current source value due to random mismatch.
ALBIOL et al.: MISMATCH AND DYNAMIC MODELING OF CURRENT SOURCES IN CURRENT-STEERING CMOS DACS 161
The aspect ratio of the CS transistors is determined by The drain-to-source impedance, considering channel length
the drain current and the overdrive voltage according to the modulation, is expressed by
MOSFET saturation drain current expression
with (7)
(2)
where is the channel length modulation parameter, the
drain-to-source saturation current, is a technology constant
The CS transistor dimensions are univocally determined once [14], is the drain-to-source voltage, and is the built-in
the overdrive voltage is chosen and vice versa, according to (1) junction potential, which is also a technology constant.
and (2). The INL of a set of theoretically equal manufactured Combining (4), (5), and (7) and expressing the
DACs is a statistical process that depends on process variations. drain-to-source voltages as function of node voltages (SW and
Systematic mismatch is addressed at the physical design phase. CAS gate bias voltage, and the output node minimum voltage
If only random mismatch is considered, the worst case INL, free , which is the worst case for output impedance) yields
of offset and gain errors, is found at the mid code in the DAC
static transfer function. In this case, the INL error for an -bit
DAC is due to the combination of LSB current sources
(independently of the DAC segmentation1 ). The LSB current
source minimum relative accuracy required to achieve a certain
INL upper bound with a given statistical yield, as a function of (8)
the DAC resolution, is [10], [13]
for the topology and
(3)
(5) (11)
Since the drain current is a value fixed by the DAC spec- C. Analysis of the Current Source Circuit Transistors
ifications, the only degree of freedom available to design the Operating Region
transconductances is either the overdrive voltage or the aspect
ratio Despite the fact that the overdrive voltage in (1) should
be maximized to minimize the CS area, it should be small
enough to allow the other transistors (SW and CAS if present)
(6) to work in saturation (the region in which higher output
impedance is obtained) in any situation.
1This is due to the fact that a source of value k LSBs is actually obtained by
1) Simple Current Cell : Applying the MOSFET
combining k sources of LSB value in parallel, instead of a single source with saturation condition to all transistors
transistors k times wider than the LSB source transistors. in the current source of Fig. 2(a) leads to the definition of a
162 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 51, NO. 1, JANUARY 2004
voltage range in which the SW gate bias voltage must be kept in random variables of the circuit that appear when mismatching
order to allow both CS and SW transistors to work in saturation effects are taking into account:
(12)
A solution exists for (12) if and only if the difference between (16)
the upper and lower bounds is positive (i.e., the lower bound for
the SW gate voltage is smaller than the upper bound
. This determines an upper bound for the addition of the
two transistors’ overdrive voltages and subsequently by calculating the partial derivatives of these
expressions with respect to each one of the random variables
(13) (that are considered independent). In this way, the variance of
the upper bound for the SW gate voltage that guarantees satu-
Resuming the saturation constraint of (12), note that the mid- ration when random mismatch effects are taken into account is
point between the upper and lower bounds for the SW gate bias found as
voltage corresponds to the optimum voltage that maximizes the
dc output impedance found in Section II-B
(14)
This is not a coincidence, since this optimum SW gate voltage
(17)
places the operating point of both SW and CS transistors at the
same voltage distance from the triode region.
If there are no other criteria for choosing the overdrive volt- Similarly, the variance of the lower bound yields
ages, such as dynamic output impedance, settling time, etc., it
would be interesting to choose the combination of overdrive
voltages that leads to the minimum area solution. This usu-
ally corresponds to an SW transistor with minimum length and
(18)
width, from which the SW overdrive voltage is derived, and the
maximum CS overdrive voltage that verifies (13), provided that
the corresponding CS transistor width and length obtained from In (17) and (18), only node voltages (except biasing gate volt-
(1) and (2) satisfy the technology minimum size constraint. By ages), current values, and load impedance have been considered
doing this, the operating points of both transistors are found just as being random variables affected by mismatching. In order to
at the limit between the triode and the saturation regions. take into account also die-to-die process variations, the worst
In the previously published current source sizing procedure case process parameter values (min or max) should be used in-
[10], which is representative of the usual design process for this stead of the typical ones. In the case of , the only param-
type of DAC, only the mismatch error of the current source tran- eter affected by die-to-die process variations is the SW threshold
sistor is considered. Hence, an arbitrary safety margin is voltage ( is always the same assuming that the full-scale
introduced as follows in the saturation constraint (13) to prevent DAC output value is adjusted to eliminate offset and gain er-
the transistors to enter triode region due to process variations: rors). Therefore, the minimum SW threshold voltage should be
used. In the case of , the variation of the SW and CS
(15) threshold voltage compensate for each other, as they appear in
(16) with opposite sign. Therefore, only the die-to-die variation
If not only the mismatch errors of the CS transistor, but also of the affects the lower bound and the worst case is the
the switches and additional cascode transistors mismatch errors minimum value for that process.
are taken into account, the overall basic current cell circuit can To find an appropriate value for the SW gate voltage, the
be optimized without introducing that arbitrary safety margin. upper bound must be larger than the lower bound in a given per-
Alternatively, this safety margin can be found by performing centage of the cases expressed by . Fig. 3 illustrates
parametric Monte Carlo simulations. The model that will be pre- this tradeoff. To accomplish that the saturation constraint is ful-
sented in the following avoids either the arbitrary design margin filled with a given probability , the optimum of the
or the time-consuming Monte Carlo simulations. SW gate voltage found in (10), which is now the mean value of
In order to include the effects of process variations in the a random variable, has to verify that
saturation condition of (12), the statistical variation of the two
bounds for the SW gate bias voltage is modeled by means of a
normal distribution. The variance of the upper and lower bounds
is found, first by expressing these bounds as a function of the (19)
ALBIOL et al.: MISMATCH AND DYNAMIC MODELING OF CURRENT SOURCES IN CURRENT-STEERING CMOS DACS 163
(20)
(21)
(22)
(24)
because the worst case of the bounds variance for the several
current sources of the DAC is found in the LSB current source which in this case of the topology lead to two
(since its area is the smallest of all the current sources), and its saturation conditions
two complementary SW transistors must be inside both of the
bounds with the same probability.
The expression of (21) represents a saturation constraint
more realistically than (15) for the CS and SW current cell
circuit, where an arbitrary safety margin was included. The (25)
safety margin appearing in (21) is not arbitrary nor needs to
be found using parametric Monte Carlo simulation. It can The equations in (25) are analytical expressions that depend
be related to other circuit parameters and mismatch process on circuit and mismatch parameters. They will be used in the
(23)
164 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 51, NO. 1, JANUARY 2004
(26)
(27) Fig. 6. Optimum as a function of poles quotient for a 12- and 10-b DAC.
Fig. 7. Optimum settling-time for the CS + SW topology. Fig. 8. Effects of the Y node in the optimum settling time for the CS+CAS+
SW topology.
differentiation, and considering that the optimum is found when
equals at the instant for which (27) is minimum:
(33)
SFDR (34)
cascoded current cell. This yields to the frequency-dependent the SW width will increase its transconductance without signif-
output impedances of (35) and (36), shown at the bottom of icantly affecting and , although the clock-feedthrough
the page. In the frequency band of interest (up to Nyquist), will be compromised.
the zeros do not affect dynamic output impedance because The previous analysis applies to the OFF–ON switching branch
for reasonable transistor sizes. of the current source. If the same criterion is considered in the
analysis of the complementary ON–OFF switching branch, the
C. Implications of the Settling Time and Dynamic Output conclusion is that during the second switching phase the time re-
Impedance in the Design Procedure sponse of the complementary output node exclusively depends
on .
In this subsection, the settling time and dynamic output Concerning the dynamic output impedance requirement, in
impedance dependencies on circuit poles and initial conditions the case of the simple current cell , a minimum-
are related to the transistors sizes and other circuit parameters. length SW transistor allows minimizing . Even if intercon-
1) Simple Current Cell : Fig. 7 depicts the nect capacitance dominates, it is not worth increasing the SW
estimated optimum settling time as a function of and transistor length to attempt to decrease the output impedance
, assuming that the optimum is obtained by the proper zero down to the pole frequency, as this compromises settling
synchronization of the switches’ control signals. From Fig. 7, time and dramatically increases the current cell area and, conse-
the intuitive rule that maximizing both poles leads to lower quently, the overall DAC area. Choosing the minimum area for
settling time is quantitatively confirmed. In order to maximize the CS transistor allows to minimize CS array area and, thus, the
, a minimum-length SW transistor should be used. The SW interconnect capacitance. Furthermore, a minimum CS width
transconductance increases as the square root of the SW tran- will also reduce node capacitance.
sistor width, but gate-to-source capacitance increases linearly, 2) Cascoded Current Cell : The same
thus a minimum-width SW transistor also maximizes . On conclusions of Section III-A1 and design implications of Sec-
the other hand, a minimum-width CS transistor minimizes the tion III-C1 are applicable to with respect to and
CS drain capacitance, therefore minimizing . A minimum . A minimum-length and -width SW transistor and min-
width for the CS implies that it must have the maximum CS imum-width CAS transistor (instead of CS) are optimum for
overdrive voltage that guarantees saturation for all transistors the settling time. This is especially true if it is taken into ac-
(Section II-C). In most cases, however, the influence of the SW count that a small interconnect capacitance exists between SW
and CS transistors’ width on can be neglected. This capac- and CAS transistors in actual circuits. In this sense, the CAS
itance is dominated by the interconnect capacitance between transistor not only increases current cell dc output impedance,
the CS and SW transistors in actual circuits, as each transistor but also improves settling time performance. Concerning the in-
is placed in a different array and their interconnections usually fluence of the additional node , as discussed in Section III-A2
are large enough to dominate the node parasitic capacitance. and from Fig. 8, it is clear that only when is low enough and
Furthermore, a minimum-width SW transistor also maximizes is high is the settling time compromised. This situation is not
by minimizing its drain and overlap capacitances. This common, since the smaller is, the smaller will be, as
is specially significant for high segmentation ratios, where a opposes voltage variations at node . In general, for reasonable
great number of switch drains are connected together at the values of , , and , and reasonable voltage variations at
output node. Apart from the benefits to the settling time, the the internal nodes, pole does not effectively affect the set-
clock-feedthrough is reduced if the SW gate and drain overlap tling time.
capacitances are lowered, which occurs when the SW width is As far the dynamic output impedance is concerned, the same
minimum. argument than in Section III-C1 may be applied to the SW and
The main design implication extracted from the previous dis- CAS transistors length. Also, minimum CS area is the best
cussion is that a minimum-length and -width SW transistor and choice for output impedance bandwidth. Finally, minimum
minimum-width CS transistor (maximum overdrive voltage) is CAS width leads to better performance, as it minimizes nodes
the optimum choice for settling time. However, the effect of SW and capacitance. Minimum transistor dimensions pro-
and CS width should be evaluated for every particular case, es- duce maximum bandwidth for output impedance and SFDR
pecially when node and parasitic capacitances are not dom- performance, apart from minimizing DAC area. Note that if the
inated by the SW capacitances. In this latter case, increasing dominant pole is , adding one or more extra
(35)
(36)
168 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 51, NO. 1, JANUARY 2004
cascode transistors between CAS and SW may improve SFDR enclosed within the boxes. The design criteria used to derive
bandwidth performance without compromising settling time, the intermediate results are shown by the arrows. Note that, in
although the overall DAC area is increased (especially if we general, the design procedures do not require the segmentation
take into account that a lower CS overdrive voltage will need ratio to be chosen previously, at least not until the driver and
to be chosen). latches are designed. The only exception is the case in which the
output node capacitance needs to be estimated with precision
IV. DESIGN PROCEDURE (this design path is indicated in Fig. 9 with an asterisk). The
proposed design procedure is closed and provides as a final
In Sections II and III, the DAC current sources’ static, dy-
result the complete sizing of the DAC current source circuits.
namic, area, and statistical models have been presented. Perfor-
mance requirements are related to the DAC design specifica-
tions, such as number of bits, linearity (INL and DNL), sam- V. CONCLUSION
pling frequency, spectral performance, yield, area, and tech- This work provides an extended design procedure for cur-
nology [1]. rent-steering DACs through the static and dynamic study of two
On the basis of the models presented in Sections II and III, a usually considered current source circuit cells. On the one hand,
current cell design procedure can be derived for both the simple an extended mismatch statistical study which takes into account
and cascoded current cell topolo- random variations in all the transistors of the current source cells
gies. yields an improved analytical static model. This refined model
Fig. 9 shows the design procedures for the two topologies. permits us to tighten the required biasing margins without re-
The starting point includes design specifications and tech- sorting to arbitrary margins, which, given the design space of
nology parameters. For the two topologies there are basically the DAC, results in notable area reductions. Conditions for op-
two targets driving the optimization process (indicated with timum gate bias voltages concerning output resistance are ob-
italics in the figure), namely minimize current cell area and tained as well. On the other hand, an extension of the switching
minimize settling time. Note that in the cascoded topology the transient behavior is discussed, and its relation to transistor pa-
minimum settling time target is coincident with the optimum rameters is discussed. As a consequence of the extended static
clock feedthrough and driver’s load target, whereas for the and dynamic modeling, an improved DAC design procedure is
simple current source topology (left part of the figure) the presented. By applying the design method to a particular tech-
settling time optimization has some negative impact on the nology and DAC specifications, comparisons with previous de-
clock-feedthrough (because of the required increase in the SW sign approaches clearly indicate that the improved modeling of
transistor width). Hence, two alternatives are shown in this case mismatching and dynamics effects presented in this work allow
depending on whether glitch energy or sampling frequency to reduce the area and improve the settling time of the DAC.
is priorized. Intermediate results of the design procedure are The results of this paper may ease the automatization of the
ALBIOL et al.: MISMATCH AND DYNAMIC MODELING OF CURRENT SOURCES IN CURRENT-STEERING CMOS DACS 169
circuit design of future current-steering DACs with improved José Luis González (M’99) received the Diploma
performance. in telecommunication engineering from the Ramon
Llull University (URL), Barcelona, Spain, in
1992 and the M.S. degree in telecommunications
REFERENCES engineering and the Ph.D. degree (with honors) in
electronic engineering, both from the Polytechnic
[1] M. Gustavsson, J. Wikner, and N. Tan, CMOS Data Converters for Com- University of Catalunya (UPC), Barcelona, in 1994
munications. Boston, MA: Kluwer, 2000.
and 1998, respectively.
[2] A. Van den Bosch, M. Steyaert, and W. Sansen, “SFDR-Bandwidth He is currently a full-time Associate Professor with
limitations for high speed high resolution current steering CMOS D/A
the Department of Electronic Engineering, Telecom-
converters,” in Proc. IEEE Int. Conf. Electronics, Circuits and Systems munication Engineering School of Barcelona, UPC.
(ICECS), Sept. 1999, pp. 1193–1196.
During 1999, he spent a semester at the Electrical and Computer Science De-
[3] P. Hendriks, “Specifying communication DAC’s,” IEEE Spectrum, pp. partment, University of Arizona, Tucson, with a Fulbright fellowship, and he
58–69, July 1997.
collaborated with the Motorola Wireless Infrastructure Group, Phoenix, AZ, in
[4] Y. Cong and R. L. Geiger, “Switching sequence optimization for gradient the investigation of noise effects on D/A converters for communications. His re-
error compensation in thermometer-decoded DAC arrays,” IEEE Trans. search interests include VLSI design, mixed-signal and RF integrated circuits,
Circuits Syst. II, vol. 47, pp. 585–595, July 2000.
and noise problems. He has authored or coauthored 26 scientific papers in jour-
[5] J. Vandenbussche, G. Van der Plas, W. Daems, A. Van den Bosch, G. nals and conference proceedings and one book chapter.
Gielen, M. Steyaert, and W. Sansen, “Systematic design of high-accu-
Dr. González received the Doctoral Excellence Award for his Ph.D. thesis
racy current-steering D/A converter macrocell for integrated VLSI sys- from the Polytechnic University of Catalunya in 2001. His thesis work has been
tems,” IEEE Trans. Circuits Syst. II, vol. 48, pp. 300–309, Mar. 2001.
published in the book Analysis and Solutions for Switching Noise Coupling in
[6] C. H. Lin and K. Bult, “A 10.b, 500-Msample/s CMOS DAC in 0.6
mm ,” IEEE J. Solid-State Circuits, vol. 33, pp. 1948–1958, Feb. 1998.
Mixed-Signal IC’s (Boston, MA: Kluwer, 1999).
[7] J. L. González and E. Alarcón, “Clock-Jitter induced distortion in high-
speed CMOS switched-current segmented digital-to-analog converters,”
in Proc. IEEE Int. Symp. Circuits and Systems (ISCAS’01), May 2001,
pp. I512–I515.
[8] H. Kohno, Y. Nakamura, A. Kondo, H. Amishiro, T. Miki, and K. Eduard Alarcón (S’96–M’01) received the M.S.
Okada, “A 350-MHz 8-bit CMOS D/A converter using delayed driving (with national honors) and Ph.D. (with honors)
scheme,” in Proc. IEEE 1995 Custom Integrated Circuits Conf., 1995, degrees in electrical engineering from the Universitat
pp. 211–214. Politecnica de Catalunya, Barcelona, Spain, in 1995
[9] G. A. M. Van der Plas, J. Vandenbussche, W. Sansen, M. S. J. Steyaert, and 2000, respectively.
and G. Gielen, “A 14-bit intrinsic accuracy Q random walk CMOS From 1995 to 1997, he was awarded a research
DAC,” IEEE J. Solid-State Circuits, vol. 34, pp. 1708–1718, Dec. 1999. fellowship from the Catalan Government. During
[10] J. Bastos, A. M. Marques, M. S. J. Steyaert, and W. Sansen, “A 12-Bit in- the period 1997–2000, he was an Assistant Professor
trinsic accuracy high-speed CMOS DAC,” IEEE J. Solid-State Circuits, with Department of Electronics Engineering,
vol. 33, pp. 1959–1969, Dec. 1998. Telecommunication Engineering School of Barcelona, where he became
[11] T. Miki, Y. Nakamura, K. Okada, and Y. Horiba, “Transient analysis of full-time Associate Professor in 2000. During the period August 2003 to
switched current sources,” IEICE Trans. Electron., vol. E75-C, no. 3, January 2004, he is holding a visiting appointment at the COPEC center,
pp. 288–295, Mar. 1992. University of Colorado, Boulder. He has participated in five Spanish national
[12] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, “Matching research projects. His current research interests include the areas of analog and
properties of MOS transistors,” IEEE J. Solid-State Circuits, vol. 24, pp. mixed-signal microelectronics with particular interest in current-mode design,
1433–1440, May 1989. theoretical aspects of fuzzy control, integrated switching dc–dc converters, and
[13] A. Van den Bosch, M. Steyaert, and W. Sansen, “An accurate statistical nonlinear controller VLSI implementations. He has authored or coauthored
yield model for CMOS current-steering D/A converters,” in Proc. IEEE more than 50 scientific papers in journals and conference proceedings and two
Int. Symp. Circuits and Systems (ISCAS), May 2000, pp. IV105–IV108. book chapters.
[14] A. J. David and K. Martin, Analog Integrated Circuit Design. New Dr. Alarcón was recipient in 1998 of the Myril B. Reed Best Paper Award at
York: Wiley, 1997. the IEEE Midwest Symposium on Circuits and Systems for a research paper on
[15] M. Albiol, J. L. González, and E. Alarcón, “Improved current-source a mixed-signal reconfigurable neuro-fuzzy processor. He is the invited co-ed-
sizing for high-speed high-accuracy current steering D/A converters,” itor of a special issue of the Analog Integrated Circuits and Signal Processing
in Proc. IEEE Int. Symp. Circuits and Systems (ISCAS), Bangkok, Thai- Journal devoted to current-mode circuit techniques and co-organizer of a spe-
land, May 2003, pp. 837–840. cial session on “New trends in switching power converters toward circuit inte-
[16] M. Albiol, “Design and implementation of a high-performance gration” at ISCAS’03.
digital–analog converter in 0.35 m CMOS technology for communi-
cations applications,” Master’s thesis, Barcelona Telecommun. School
(ETSETB), Univ. Politècnica de Catalunya (UPC), 2002.
K.U.Leuven, ESAT-MICAS
Kasteelpark Arenberg 10, B-3001 Leuven, Belgium
e-mail: tao.chen@esat.kuleuven.ac.be
ABSTRACT Vb
Vb
A behavior-level model of a thermometric DAC with acceptable Ro
complexity and sufficient accuracy is presented in this paper. The Vo (t) Ro
impact of the current sources’ internal poles on the SFDR is de- Vo (t)
Vg2(t) Vg3(t)
scribed in the model. With this model, the sensitivity of the SFDR Co
to some unideal factors such as the parasitic capacitors in the inter- Vx (t)
nal point, and the variation of the internal points’ voltages, can be Internal node X Vx (t)
observed directly. Furthermore, this model can be used for the ver- Vg0 +
Ŧ
In the previous model, the transition behavior of the DAC is con- −I2 + gm2 [V1 (t) − V 1F ] − go2 [Vo (t) − V1 (t)]+
tained in the two parameters vo0 and vx0 . The output conductances
d
of the transistors are approximated to zero in this model. So this I1 + go1 V1 (t) + C1 V1 (t) = 0 (8)
is not enough to predict the impact of the output impedance on the dt
DAC’s SFDR, which has been proved to be one of the main factors −I4 + gm4 [V2 (t) − V 2F ] − go4 [Vo (t) − V2 (t)]+
that limit the SFDR of a DAC [1]. We will extend the mathemat- d
I3 + go3 V2 (t) + C2 V2 (t) = 0 (9)
ical model to take the output resistances into consideration in this dt
section, and then in section 3 we will get the behavioral model
based on the results obtained. It’s convenient to express the node voltages by using the devi-
The model used is shown in Fig. 2. The left box in the figure ations from their final values VoF , V1F and V2F as variables:
is the model of the current sources which are switching on during
the current sampling cycle, and the right box is the model of the vo (t) = Vo (t) − VoF (10)
current sources which were already switched on at the beginning v1 (t) = V1 (t) − V1F (11)
of the current sampling cycle. Those current sources which remain v2 (t) = V2 (t) − V2F (12)
off are neglected since their transconductances are very small, even
when the signal frequency is high. And notice that the final values of the voltages satisfy the fol-
lowing equations:
Vb
I2 (t) = I2 − gm2 [V1 (t) − V1F ] (5) 2.3. Some Special Cases
I4 (t) = I4 − gm4 [V2 (t) − V2F ] (6)
Some special boundary cases still need to be considered.
Applying these results, the node equation with respect to the When all the current sources are switched off at the beginning
output node is: of the current cycle, the circuits in the right box of Fig. 2 don’t
exist. The circuits equations then become:
go [Vo (t) − Vb ] + I2 − gm2 [V1 (t) − V1F ] + go2 [Vo (t) − V1 (t)]+
(go + go2 + sCo )Vo (s) + (−gm2 − go2 )V1 (s) = Co Vo0 (22)
d
I4 − gm4 [V2 (t) − V2F ] + go4 [Vo (t) − V2 (t)] + Co Vo (t) = 0 (7) −go2 Vo (s) + (gm2 + go2 + go1 + sC1 )V1 (s) = C1 V10 (23)
dt
,
When no current source is switched on during the current cy- With equations (28) and (29), the initial value of vo (t) can be
cle, the circuits in the left box of Fig. 2 don’t exist. The circuit obtained:
equations are then: Vo0 = Vo (0) − Vdif f − Vo (∞) (31)
where Vdif f is the difference between the voltage at the end of the
(go + go4 + sCo )Vo (s) + (−gm4 − go4 )V2 (s) = Co Vo0 (24) last cycle and the ideal voltage when time is approaching infinity.
−go4 Vo (s) + (gm4 + go4 + go3 + sC2 )V2 (s) = C2 V20 (25) It describes the fact that the ideal final value can never be reached.
The value of V10 is decided by the crossing point of the two
When no current source is switched on at both the beginning control signals (Vg2 (t) and Vg3 (t) in Fig. 1(a)), and is control-
and the end of the current cycle, the circuits can be simplified into lable.
an R-C network, and: For the current sources which are always on during this cycle,
the deviation of their internal point (point B in Fig. 2) is caused by
Vo0 the deviation of the output point. Since Ro 1/go3 , we can get:
Vo (s) = (26)
s + go /Co
Vo0
V20 = (32)
1 + gm4 /(go3 + go4 )
2.4. The Initial Conditions
In order to obtain the output voltage vo (t), the initial values Vo0 , 3. MODELING WORK
V10 and V20 have to be calculated. These initial values are the de-
viations of the voltages at the beginning of the current cycle from With the mathematical model described in section 2, a behavioral
their final values. model of the DAC can be obtained. Fig. 4 shows the flow chart
At the beginning of the current cycle, only the current sources of the program. In our program, all the parameters are extracted
in Fig. 2’s right box are on. The circuits for calculating the voltage from the simulation results of a real design[3], except the parasitic
at the output node are shown in Fig. 3(a), where the cascode cur- capacitance in the internal node. For simplification of the model,
rent source is simplified into an ideal current source and an output the simulation results below are based on the assumption that the
resistance ro34 : parasitic capacitance at the internal node of every current source
is the same. The total parasitic capacitance depends on the area of
ro34 = 1/go3 + 1/go4 + gm4 /(go3 go4 ) (27) the current sources block and can be estimated. In our case, the
estimated value is about 40 pF.
The output voltage at the beginning of the cycle can be obtained
as:
ro34 ro34 Ro For every sampling cycle,
Vo (0) = Vb − I3 (28) calculate the number of current
Ro + ro34 ro34 + Ro
sources that have been on and
that will be switched on
Figure 3: The schematics for the calculation of the initial values of Figure 4: The flowchart of the behavioral model
the output voltage
Fig. 5(a) shows the output signal of the DAC. Fig. 5(b) is the
At the end of the cycle, the current sources in both boxes of result of the FFT analysis. The SFDR can be obtained by calculat-
Fig. 2 are on. The circuits are shown in Fig. 3(b). The output ing the difference between the amplitude of the signal frequency
voltage is: and the amplitude of the second largest value in the frequency
spectrum.
ro34 ro12 (ro34 ro12 )Ro Doing the same calculation for different signal frequencies, we
Vo (∞) = Vb − (I1 + I3 ) (29)
Ro + ro34 ro12 r034 ro12 + Ro get the SFDR-fsig curve as shown in Fig. 6(a). We see from this
figure that when the signal frequency is higher than a value (which
where is quite low), the SFDR will decrease with increasing signal fre-
ro12 = 1/go1 + 1/go2 + gm2 /(go1 go2 ) (30) quency at a slope of about -20dB/dec, until it reaches a relative
,
a real converter, the SFDR might not really decrease. But still we
The Output of the DAC The Power Spectrum Deisity of the Output Signal
0.3 40
can expect that the SFDR won’t increase a lot by optimizing the
0.2
20
0
crossing point of the control signals.
0.1
fsig=9MHz Ŧ20
Another interesting relation is the relation between the SFDR
Output Voltage (v)
Ŧ100
the SFDR will decrease with increasing total capacitance value
Ŧ0.2
Ŧ120 at a rate of -20dB/dec. This is because the pole of the output
Ŧ0.3 Ŧ140
impedance is reduced when the value of the parasitic capacitor is
Ŧ160
0 2 4 6
Time (s)
8 10 12
Ŧ8
x 10
0 1 2 3 4
Frequency (KHz)
5 6 7 8 9 10
4
x 10 increasing, thus the SFDR will be reduced to lower value at the
given signal frequency.
(a) The time-domain output (b) The frequency-domain
signal of the DAC output signal of the DAC 4. CONCLUSIONS
82
[5] [6].
The variation of the internal points’ voltages is also included
105
80
100 fsig=29MHz
78 in the model, and their impact on SFDR is also investigated. The
95
results show that the SFDR cannot improve a lot by adjusting this
SFDR (dB)
SFDR: DB
76
Ctot=40pF
90
85 V =Ŧ0.0465V
74
parameter.
10
72
80 C =40pF
tot 70
75 68 5. REFERENCES
70 66
10
3 4
10
5
10 10
6 7
10 10
8 Ŧ1 Ŧ0.8 Ŧ0.6 Ŧ0.4 Ŧ0.2 0 0.2 0.4 0.6 0.8 1
85
f =29MHz
sig [3] G.Van der Plas, J.Vandenbussche, W.Sansen, M.Steyaert
80
v =Ŧ0.0465V
10
and G.Gielen, “A 14-bit Intrinsic Accuracy Q2 Random
Walk CMOS DAC, ” IEEE J. Solid-State Circuits, vol. 34,
75
,
Modelling and implementation of a
10-bit 80 MSPS current-steering DAC
with internal bandgap reference
in a 0.18 µm CMOS process
Department of Electroscience
Abstract
A 10-bit quadrature differential digital to analog converter has been designed in a 0.18
µm CMOS process. This DAC is a part of the SoCTRix project at Acreo. There has
been a large effort put into MATLAB modelling of DAC behaviour and errors. A
design flow was built up to systemize and simplify the layout work. To minimize
external components, a bandgap was implemented in the same design. The main
design goal was to reach high linearity while keeping power consumption low. A
secondary goal has been to keep the total chip area low. In an effort to make the DAC
commercially viable it has been designed to give a high yield.
1
2
Preface
This thesis work has been carried out at Acreo AB in Lund. The thesis is part of our
Master of Science degree from Lund Institute of Technology – LTH, Lund University.
We would like to thank our advisors Karl Norling from Acreo, Martin Anderson from
Department of Electroscience, LTH and all co-workers at Acreo.
We would also like to thank our examiner at LTH, Jiren Yuan.
3
4
1 Abbreviations........................................................................ 7
2 Introduction .......................................................................... 9
2.1 Design goals...................................................................................................9
3 Theory................................................................................. 11
3.1 DAC Theory.................................................................................................11
3.2 Performance measures .................................................................................11
3.2.1 Offset error...........................................................................................11
3.2.2 Gain error .............................................................................................12
3.2.3 Differential Nonlinearity......................................................................12
3.2.4 Integral Nonlinearity............................................................................13
3.2.5 Monotonicity........................................................................................14
3.2.6 Glitches ................................................................................................14
3.2.7 Settling time .........................................................................................14
3.2.8 Signal-to-Noise and Distortion Ratio...................................................14
3.2.9 Spurious Free Dynamic Range ............................................................14
3.2.10 Resolution ............................................................................................15
3.2.11 Accuracy ..............................................................................................15
3.2.12 Effective Number Of Bits ....................................................................15
4 DAC Architecture............................................................... 17
4.1 Choice of architecture ..................................................................................17
4.2 Current-Steering DACs................................................................................17
4.2.1 Thermometer code ...............................................................................18
4.2.2 Thermometer-Binary segmentation .....................................................18
5 Modelling of errors in a current-steering DAC................... 21
5.1 Causes of errors in current-steering DACs ..................................................21
5.1.1 Output impedance vs. SFDR................................................................21
5.1.2 Output impedance vs. gain error..........................................................21
5.1.3 Mismatch vs. SFDR .............................................................................22
5.1.4 Mismatch vs. INL ................................................................................22
5.1.5 Influence of noise on circuit performance ...........................................23
5.1.6 Influence of graded errors on performance..........................................23
5.1.7 Mismatch calculations .........................................................................24
5.2 MATLAB model..........................................................................................25
5.3 MATLAB simulations .................................................................................25
5.3.1 Simulating the nominal case ................................................................25
5.3.2 Simulating the impact of reduced matching ........................................27
5.3.3 Simulating the impact of switching schemes.......................................28
5.3.4 Simulating the impact of limited output impedance ............................30
6 System implementation ...................................................... 31
6.1 System description .......................................................................................31
6.2 Implementation of sub-circuits ....................................................................32
6.2.1 Current source......................................................................................32
6.2.2 The binary-to-thermometer decoder ....................................................35
6.2.3 Latches .................................................................................................37
6.2.4 Bandgap reference ...............................................................................40
6.2.5 Voltage-to-current converter................................................................45
6.2.6 Voltage-to-current converter bias circuitry..........................................48
5
7 Results and conclusion ....................................................... 49
7.1 The complete system....................................................................................49
7.2 Challenges of system simulations................................................................49
7.3 Simulation results.........................................................................................50
7.4 Conclusions..................................................................................................50
7.5 Possible improvements for future development ..........................................51
8 Measurements ..................................................................... 53
8.1 Introduction..................................................................................................53
8.2 Static measurements.....................................................................................53
8.3 Dynamic measurements ...............................................................................54
9 References .......................................................................... 57
Appendix A............................................................................... 59
Appendix B............................................................................... 63
Appendix C............................................................................... 69
6
1 Abbreviations
Abbreviation Explanation
OP Operational Amplifier
SE Silicon Ensemble
7
8
2 Introduction
Wireless communication systems today require more and more digital signal
processing. But somewhere along the line of a transmitter or a receiver there is a radio
signal, an analog signal. This requires the use of a digital-to-analog converter in a
transmitter or an analog-to-digital converter in a receiver. These communication
systems are often mobile and thus require all of the circuits involved to have low
power consumption.
9
10
3 Theory
where the vector b is the binary word and N is its length (the DAC’s resolution). The
minimum voltage change on the output, corresponding to a change of 1 LSB, is
Vref
VLSB = ( 3.2 )
2N
The maximum output signal can be calculated from equation ( 3.1 ). The decimal
value is
N
∑2
k =1
−k
= 1 − 2−N ( 3.3 )
11
Vout (B )
Eoffset ≡ ( 3.5 )
VLSB B = 0
The offset and gain errors are described graphically in Figure 3.2.
The vector’s maximum is often used in specifications to show the worst-case error.
12
Figure 3.3 Transfer curve with errors on the left, and the DNL vector on the right.
Figure 3.4 INL errors compared to a straight line between start and end points.
The equation for the INL vector is
13
Vout (B ) − Vout _ fit (B )
INLB ≡ ( 3.8 )
VLSB 0 ≤ B ≤ 2 N −1
where Vout_fit is the straight line from zero to full-scale. If the DNL and INL formulas
are combined we get an expression for DNL and INL.
DNL B = INLB − INLB −1 ( 3.9 )
B
INLB = INL0 + ∑ DNLk ( 3.10 )
k =1
3.2.5 Monotonicity
A DAC has monotonic behaviour if the output always increases with an increasing
input. Monotonicity is guaranteed if the maximum INL error is smaller than 0.5 LSB
and the DNL error less than 1 LSB. The non-monotonicity often occur when the input
is changed from “01…1” to “10…0” in binary weighted DAC’s.
3.2.6 Glitches
If there is skew between the input bits or the clock, glitches can occur. For example, if
an input word is changed from “1000” to “0111” and the MSB is slower than the
other bits. Then the input is “1111” before it changes to “0111”. The LSB step will
begin with a peak and then settle to the desired voltage.
14
Figure 3.5 Frequency spectrum showing SFDR.
3.2.10 Resolution
The resolution is defined as the number of analog levels. Which is 2N for a converter
with N physical input bits. However, this does not say anything about the actual
performance of the DAC.
3.2.11 Accuracy
The absolute accuracy of the converter is the difference between the ideal and the
actual performances, where the offset, gain and nonlinearity errors are included.
Relative accuracy is the absolute accuracy without the gain and offset errors.
15
16
4 DAC Architecture
• Pipelined
• R-2R ladder
• Resistor string
• Current steering
The pipelined converter would need to use 10 fairly fast operational amplifiers for 10-
bit operation. These would consume a lot of power to operate at the speeds required.
The output from the pipelined converter is voltage and would therefore have to be
converted into current in order for the next stage to work.
The R-2R and resistor string architecture have the same basic flaw. It does not make
efficient use of the power. This is caused by the fact that here is always a current
flowing through the resistive network. It is also hard to match the resistors for 10-bit
accuracy without a significant penalty in either area or power.
This leaves only the current steering architecture, which is chosen because it is
possible to reach 10-bit accuracy at the required speed.
Figure 4.1 A 1-bit current-steering DAC with current source and digital input b.
17
The power losses for this type of converter come from generating the bias and
synchronization of the input signal. This also means that the power consumption will
scale fairly well with the update frequency since the power consumption in a digital
net is proportional to the frequency.
A way, different from the binary, to represent a digital word is with thermometer
code.
Decimal Binary Thermometer
0 000 0000000
1 001 0000001
2 010 0000011
3 011 0000111
4 100 0001111
5 101 0011111
6 110 0111111
7 111 1111111
Table 4.1 Decimal numbers coded with binary and thermometer code.
The number of bits required for the thermometer code is
NT = 2 N B − 1 ( 4.2 )
where NB is the number of binary bits. This results in a larger amount of hardware
when there is a large word-length. However, the thermometer coded D/A converter
also has positive qualities. For example, if a binary number changes from “100” to
“011” there may be a glitch. If the same change is done with the thermometer code it
just changes one bit, and this glitch problem is gone and the DNL errors are reduced.
If the converters current sources have different current, the “011” level might give
more current than the “100” level, which means that the converter is non monotonic.
This is impossible with a fully thermometer coded current-steering converter because
the number of active current sources always increases with the input, and this
guarantees monotonicity.
18
Thermometer
Positive
• Low glitch energy
• Monotonicity
• Small DNL errors
Negative
• Digital decoding with more area and power consumption
• Increased number of control signals
Binary
Positive
• Low digital power consumption
• Smaller area
• Small number of control signals
Negative
• Monotonicity not guaranteed
• Larger DNL errors
The price in digital power and area was rather small compared to the benefits of
smaller glitch energy and DNL errors, so a split with 8 thermometer and 2 binary bits
was chosen for this DAC. This results in a thermometer unit current, It, equal to four
binary units, Ib. One can either represent the thermometer unit as 1 current source with
4 times the binary unit current, 2 sources with 2 times the unit current or 4 unit
sources. With 1 source, the standard deviation for the transistor mismatch on a wafer
is
N ∈ (4m,4σ ) ( 4.3 )
and with 4 sources of the same size as above
( )
N ∈ 4m, 4σ 2 = (4m,2σ ) ( 4.4 )
Therefore it is better to use 4 thermometer sources because the required matching is
decreased, and it is easier to apply different layout techniques such as common-
centriod. It is also easier to match the sources because they are the same size as the
binary sources and they will have the same bias.
19
20
5 Modelling of errors in a current-steering DAC
In this chapter the cause of some of the errors causing performance degradation will
be discussed. Some approximate formulas for estimating performance are shown.
Both static and dynamic errors are taken into consideration. A MATLAB model of a
current steering DAC is shown. Some simulation results from this model are
presented.
By doing a taylor series expansion and identifying the harmonics it’s possible to
decide the influence of Zimp on SFDR. The series expansion was done in maple for
both the single-ended case and the differential case. In Appendix A the complete
maple code can be seen for both cases. In the single-ended case the SFDR is
dependant on the second order harmonic, which is the dominant overtone. In the
differential case, assuming perfect matching, the second order harmonic is cancelled
and the SFDR is dominated by the third overtone.
The ratio between the signal and the second order harmonic in the single-ended case
is, according to [3], approximately given by:
N
Z imp
SFDR = ( 5.2 )
4 2N
+
RL Z imp
For a 10-bit converter SFDR should be at least 60 dB. In this case with RL=60 Ω this
would mean that Zimp would have to be at least 15.3 MΩ in the desired frequency
range. In the differential case an output impedance of 455 kΩ is enough to meet the
demands. A demand that is much easier to meet, especially if high frequency
operation is desired.
21
current sources are on and the value of the load resistor. The worst case occurs when
all current sources are conducting. The full-scale gain error is easily determined by
the following equation:
Vout Z imp
= ( 5.3 )
Vin 1023RL + Z imp
The only problem with this simple equation is that it doesn’t take the frequency
dependence of Zimp into consideration. Since the frequency behaviour of Zimp is hard
to determine analytically one has to rely on simulations for this.
where σI/I is the standard deviation of a unit current source, N is the number of bits,
inv_norm is the inverse cumulative normal distribution and yield specifies the
percentage of DACs with an INL below ½ LSB. Figure 5.1 shows the yield plotted
versus the current source mismatch.
22
In a 10-bit DAC this means that the matching has to be around 0.5% to get a
reasonable yield.
23
Figure 5.2 Illustration of the hierarchical symmetric switching scheme.
First current source 1 is switched on and then when number 2 is switched on it cancels
out the symmetrical error. Number 3 and 4 cancel their symmetrical error in the same
way and together with 1 and 2 they cancel the graded error as well.
As described earlier each thermometer coded current source is divided into four
separate current sources. This means that there will be four arrays of 255 current
sources. Each one of these arrays has their current sources placed according to the
switching scheme above. To further reduce the impact of gradients a common-
centroid layout scheme is used, meaning that all current sources will have a common
centre. This is accomplished by mirroring the four arrays about the centre. This
method works well to cancel linear gradients but does less to cancel parabolic
gradients.
24
5.2 MATLAB model
To study the impact of all the errors above together a MATLAB model was created.
The model consists of three files DAC.m which is the model of the DAC, swmatris.m
which generates the matrix for the switching scheme and run_DAC.m in which
simulation parameters are entered. The code for these files can be seen in Appendix
B. At first an ideal ADC is used to create the input signal to the DAC. Some noise is
added to the input signal of the ADC to avoid spurious tones in the final spectrum.
Parameters like those described above are fed into the model. These include noise,
gradients and output impedance for instance. A current source matrix is created in
MATLAB using the data on gradients and random errors (mismatch). An example of
how this matrix may look can be seen in Figure 5.3 below.
25
Figure 5.4 Output spectrum.
26
5.3.2 Simulating the impact of reduced matching
In the next following figures the impact of increased mismatch is shown by changing
current source matching from 0.5% to 1.5%.
27
Figure 5.12 and Figure 5.13 showing INL and DNL shows more of a deterioration in
performance when mismatch increases. Not very many of the DACs are below the
wanted level of 0.5 LSB INL. At 1.5 % mismatch the equation ( 5.5 ) shown earlier
predicts that only 70 % of the DACs will be below 0.5 LSB INL. The simulations
suggest that it’s even fewer than that. This is most probably caused by the gradients
which aren’t taken into account in equation ( 5.5 ).
28
Figure 5.15 SFDR. Figure 5.16 ENOB.
As can be seen from the spectrum in Figure 5.14 the second harmonic overtone
becomes dominant and other spurs have also increased in amplitude. This is caused by
gradients that are not cancelled well in the conventional switching scheme. Since a
common-centroid layout is used the linear gradients are perfectly cancelled, the
parabolic gradients however, are not. From Figure 5.15 showing SFDR it can be seen
that the mean value has gone from around 72.5 dB to around 56 dB. That equates to a
loss of 16.5 dB in SFDR.
Looking at Figure 5.16, showing the ENOB, a significant loss in performance is seen
here as well. It has gone from 9.8 bits to 8.8 bits, which equates to a loss of 6 dB in
SNDR so it is not as bad as the SFDR.
In Figure 5.17 and Figure 5.18 where INL and DNL are shown it can be seen that the
INL has gone from around 0.5 LSBs to 2.9 LSBs. This is also caused by the fact that
gradients aren’t cancelled well in the conventional switching scheme. The DNL
however is not any different with this switching scheme. This is easily understood
since DNL is the difference between two adjacent current sources. This means that the
error is not that big since the distance between two sources is fairly small which
means that the change in current caused by gradients will be fairly small as well.
29
5.3.4 Simulating the impact of limited output impedance
Another interesting simulation that can be done is testing the impact of output
impedance. By lowering the output impedance in the model to 455 kΩ and looking at
the output spectrum. Figure 5.19 below shows the spectrum from this simulation.
30
6 System implementation
31
the voltage in that node. Since both input and output voltages from the voltage-to-
current converter are measurable the offset voltage in it can be calculated.
6.2.1.1 Theory
In its simplest form a current source for a differential current-steering DAC might
look like in Figure 6.2 below.
32
Figure 6.3 A cascoded current source.
In order to bias this current source and get the maximum possible voltage swing
available on the output, a wide-swing architecture is used [1]. It works by biasing
PM0 so that its drain-source voltage is as close to the minimum voltage possible
without going into the triode region. In order to achieve this the bias circuit shown in
Figure 6.4 is used.
6.2.1.2 Implementation
The size of the current source transistor is determined by mismatch calculations. It has
to have a certain area to achieve the required matching as described in section 5.1.7.
The size of the cascode is determined by simulations to achieve maximum output
impedance. The cascode will also have some impact on matching but the exact
numbers are hard to determine. The switches are made as small as possible to reduce
glitches on the output. One difference from the theory above is the fact that there are
four current sources connected to each switch for layout technical reasons.
When it comes to the bias circuitry there are some differences from the theory above.
Since there are only four bias circuits per DAC, because of area and power
constraints, each bias circuit is loaded by 255 current sources. To keep the reference
voltages stable the current in each bias circuit is 8 times the current in one current
source. This lowers the output impedance of the bias circuit. To further help keeping
the voltage stable both Vb1 and Vb2 are decoupled to VDD_A. The reason for
connecting the decoupling capacitor to VDD instead of GND is that the output current
is determined by Vgs-PM0, which is referred to VDD.
6.2.1.3 Results
Shown in Figure 6.5 is the result from a simulation on output impedance. In order to
achieve 10-bit performance the output impedance in one current source has to be
larger than 455 kΩ. The smallest unit is four current sources connected through one
switch and that is what has been simulated. This means that in the figure below the
33
value has to be above 113,75 kΩ in order to achieve 10-bit performance. This is true
for all frequencies below 635 MHz. Which means that theoretically this is the limit on
the analog bandwidth.
This limited output impedance causes a gain error at higher frequencies as described
earlier. In Table 6.1 the gain error in % can be seen versus frequency.
Another important parameter when it comes to the current source is the amount of
noise that one current source produce. As describe earlier in section 5.1.5 the noise
level has to be lower than the quantization noise. Simulations show that one current
source has about 303 nV RMS noise. Using equation ( 5.8 ) one can calculate the
SNR of the DAC. The load resistor is 60 Ω and differentially this equates to a voltage
noise of 8.9 µV RMS with a bandwidth of 40 MHz. The quantization noise has a
value of 169 µV RMS with a 600 mV full-scale voltage. Putting these numbers into
the equation for SNR yields a value of 61.94 dB, which is very close to the theoretical
maximum of 61.96 dB.
1 0.02
8.125 0.2
16 0.38
32 0.77
64 1.53
128 2.95
256 5.66
34
6.2.2 The binary-to-thermometer decoder
6.2.2.1 Theory
A systematic approach of the binary to thermometer decoder is based on a 2-bit
converter [7]. To increase the number of binary inputs, a new logic level is added for
each extra bit. Truth tables and De Morgan’s laws are used between every stage to
find the optimum number of gates.
( 6.1 ) ( 6.2 )
This method makes it easy to automatically generate a schematic with the desired
number of inputs.
Figure 6.7 Even respectively odd binary number of inputs at schematic level.
35
6.2.2.2 Design
A SKILL script was written to generate the decoder schematic. The script can be
found in Appendix C. When adding an input bit, it gets a large fan-out, 2 N − 1 . In this
8-bit converter the binary MSB’s were buffered to compensate for the increasing rise
and fall time.
Verilog-XL produces a verilog netlist when a simulation is executed, and that netlist
was used to import the design into Silicon Ensemble.
The matrix signals must be rearranged somewhere between the Flip-Flops and
switches to fit our switching scheme. To facilitate the layout work, the output pins
placements were changed in SE.
By combining GDS and DEF-exports from SE, the design was imported back to
Cadence Virtuoso for DRC, extraction, LVS, simulation and some post route
optimization [8].
6.2.2.3 Results
The simulated maximum delay through the decoder was 645 ps on a schematic level,
which is more than enough with our clock strategy. Since there is a delay of 50 ps
between the clock to the flip-flops and the latches the actual delay will be 695 ps. This
delay sets the DAC’s highest update rate to 720 MHz. The extra delay is there to
make sure that the latch is in its non-transparent mode when new data is clocked in.
The flip-flop is triggered on the positive edge and the latch is in its transparent mode
when the clock is high and in its latched mode when the clock is low.
36
Figure 6.9 The decoder’s maximum tolerated delay at 80 MSPS.
6.2.3 Latches
6.2.3.1 Theory
A matrix of latches is used to synchronize all thermometer-coded bits with the binary
ones. The output bits are differential and connected to the switch array. The outputs
are designed to switch on one side before it switches off the other because we want to
keep the current sources from turning off during switching. The outputs intersect just
above zero volts since the switches are made with PMOS transistors.
6.2.3.2 Design
The latch is based on a modified clocked SR-flip-flop [9]. The set and reset inputs are
replaced with data and inverted data, D and DN in Figure 6.11 below.
37
Figure 6.11 Schematic of the latch.
The first inverters have slow N and fast P-transistors to delay signals with rising
edges, and the last inverters buffers the outputs. NM8 and NM9, the clock transistors,
are minimum size to reduce clock feed through. The buffers also contribute to this.
The current source matrix layout sets the switch array height, which is equal to the
height of the latches to simplify their connections. There are 32 rows of latches in the
array, the same as the switches and current sources. Since there are 8 thermometer and
2 binary bits, it results in 8 latches per row, and one row with an extra latch. To
prevent skew between the outputs of this large block, a clock tree was implemented.
6.2.3.3 Results
The output delays when receiving a positive edge from the clock tree are shown in
Table 6.2. The arrows stand for rising or falling edges of the signal.
clk ↑ t [ps]
D↑ 225
DN ↓ 163
D↓ 216
DN ↑ 291
38
The switch is on when Vg drops below 180 mV, which is the maximum required
intersection point. This was achieved at all corners. See Table 6.3.
0 6 4 30 4 9
1.65
110 17 11 68 7 26
0 9 6 44 5 14
1.8 27 12 8 52 6 17
110 25 17 86 11 36
0 13 9 56 6 19
1.9
110 30 22 100 13 44
Table 6.3 The intersection points at all process and environmental corners.
There is clock feed through when data is stored on the latch. The first inverter reduces
this spike by 23.4 dB. And the buffer decreases it with 17.3 dB.
39
6.2.4 Bandgap reference
6.2.4.1 Theory
The basic idea behind a bandgap reference is to create a stable voltage reference
independent of process variations and temperature. The basic principle of the bandgap
is to use a forward biased diode (base-emitter junction) as a base for the reference
voltage. Because the voltage over the base-emitter junction has a negative temperature
coefficient one has to compensate for this by adding a voltage that is proportional to
absolute temperature (PTAT). The PTAT is realized by amplifying the voltage
difference between two forward biased junctions biased at different current densities.
Figure 6.12 below shows the conventional bandgap architecture in a CMOS process.
Although independent bipolar transistors are not available in CMOS, well transistors
with their collector connected to the substrate (ground in a p- substrate) usually are.
Vref −0 = VG 0 + (m − 1)
kT0
( 6.4 )
q
where VG0 is the bandgap voltage of silicon at 0 K (equal to 1.206 V).
In the case where T0 = 300 K this implies that:
Vref −0 = 1.24V ( 6.5 )
for zero temperature dependence. Which is basically the bandgap voltage plus a
correction term for second order effects, hence the name bandgap reference for this
type of reference.
The problem with the conventional bandgap is that in a 0.18 µm process the 1.24 V
output voltage referred to ground leaves little headroom to the supply voltage, which
normally sits at 1.8 V. In the DAC, the bandgap reference voltage is fed to a voltage-
40
to-current converter, which creates the reference current for the current sources. Since
getting this converter to work with an input voltage that close to the supply voltage is
fairly hard it is desirable to have an output voltage from the reference that is closer to
ground.
For reasons stated above the architecture chosen for the bandgap is a low-voltage one
[10]. The advantage with this kind of architecture is that it gives a larger margin to the
supply voltage. Another nice feature of this architecture is the fact that the output
voltage is easily changed.
6.2.4.2 Implementation
The low voltage architecture implemented is shown below in Figure 6.13. The basic
principle of this architecture is the same as for the conventional one. It uses two
junctions biased at different current densities. Since PM1 and PM3 are biased equally
and are of the same size the current through Q0 and Q1 is identical. To get a different
current density through Q0 and Q1 transistor Q1 is made larger than Q0. Practically,
this is solved by paralleling a number of identical devices. The output voltage from
the bandgap can, according to [10], be written as:
⎛V dV ⎞
Vref = R3 ⎜⎜ f 0 + f ⎟⎟ ( 6.6 )
⎝ R2 R5 ⎠
where Vf0 is the voltage over Q0 and dVf is the voltage over R5. For a more detailed
description of the bandgap see [10]. As seen from ( 6.6 ) the output voltage is only
dependent on the ratio between the resistors and not on the absolute value.
41
With this type of circuit in an on-chip environment it’s hard to analytically calculate
the values of the resistors. This is because the resistor values are also dependent on
temperature. To solve this problem the optimizer function in cadence was utilized. It
works by giving it a number of variables to sweep and a few demands to meet. In this
particular case the values of R4 and R2 were set and R5 and R3 were swept. The
demands to meet were set as an absolute value as close to 500 mV as possible and
Vref-max –Vref-min lower than 300 µV from 0 to 110º C.
During the layout phase a few things were taken into consideration. Since the ratios
between the resistors sets the output voltage and also affects the temperature stability
it’s important that the resistors are well matched. To assure adequate matching the
resistors were divided into unit blocks of similar size. These unit blocks are then
distributed evenly over the total area that the array of resistors consumes. To ensure
that the matching is adequate after all the unit blocks are connected together one has
to consider the impact of parasitic resistance. The parasitic resistance comes from
interconnect wires and contacts. The resistor array has been laid out so that the
relative error caused by parasitic resistance is the same for all resistors.
The bipolar transistors also have to be well matched. To ensure good matching a
common centroid layout scheme was used. By placing Q0, which is a single
transistor, in the middle and the individual blocks that make up Q1 around it a square
matrix with a common centre for both Q0 and Q1 was created. An outer ring of
dummy transistors ensures that all transistors have an equal surrounding environment.
6.2.4.3 Results
Simulations were run over the specified temperature range of 0-110 ºC to verify its
function. In Figure 6.14 below the output voltage vs. temperature can be seen in the
typical case with a supply voltage of 1.8 V. Simulations were also run in the different
process corners and different supply voltages to make sure that the bandgap operates
properly under worst case conditions. The results of these simulations can be seen in
Table 6.4 below.
For future reference it was also interesting to see how low the supply voltage could go
before the output strays to far from the nominal value. The result of this can be seen in
Figure 6.15 below. The conclusions that can be drawn from this figure is that the
bandgap works very well down to around 1.2 V supply voltage.
The influence of power supply noise was investigated by looking at the power supply
rejection ratio, PSRR. The simulation was run with an external capacitor of 470 pF
through a model of the bond wire. A plot of PSRR versus frequency can be seen in
Figure 6.16. The PSRR is good at low and medium frequencies. At low frequencies
because the opamp gain is high there and the feedback loop will eliminate a lot of the
noise. At medium frequencies the capacitor on the output will attenuate the noise. At
high frequencies it’s very hard to eliminate the noise because the external capacitor’s
effectiveness is limited by the bond wire.
Finally a test was made to see how long it takes for the output voltage to stabilize
after power on. The simulation was done with an off-chip capacitor with a value of
470 pF. This value was chosen because up to this value the increase in PSRR is
notable. Any larger than this provides very little gain in PSRR and only increases
startup time. The result is shown in Figure 6.17 below. The first part of the curve is
where the startup circuit is still on and provides extra current to charge the external
capacitor. Around 350 mV the startup circuit deactivates and the capacitor is charged
with a constant current.
42
Figure 6.14 Reference output voltage versus temperature.
Table 6.4 Output voltage and relative error in different process corners.
43
Figure 6.15 Output voltage versus supply voltage.
44
6.2.5 Voltage-to-current converter
6.2.5.1 Theory
Since the reference has a voltage output and a current reference is needed for the
current sources a design is needed to convert between the two. Figure 6.18 below
shows the basic concept behind the converter [11].
6.2.5.2 Implementation
In this design the opamp was realized with a folded-cascode architecture [11]. The
advantage of this is that the converter is fully functional down to 0 V input voltage. In
Figure 6.19 the schematic of the converter is shown.
45
Figure 6.19 Schematic of the complete voltage-to-current converter
The purpose of the converter is, as stated, to generate the reference currents for the
DAC. Each DAC needs 8 reference currents since the bias is distributed over the chip.
Since there are two DACs the total number of currents needed is 16. In Figure 6.19,
only three of these are shown. To generate all these currents the output current from
the converter is mirrored through PM11 and PM12 down to NM20, which forms a
current mirror with the 16 output transistors.
The capacitor on the output of the opamp is there to stabilize the circuit. Since Vout is
fed to an off-chip resistor it will have to pass through a bond-wire. This will make the
whole circuit susceptible to instability by decreasing the phase margin. It is therefore
important to make the capacitor slightly larger than it would be with an on-chip
resistor.
6.2.5.3 Results
Important things to check for in a design like this are stability, linearity and
temperature stability. In Figure 6.20 below the AC response is shown. Phase margin
in this figure is around 82º. With a model of the bond wire and external parasitic
capacitance connected to Vout, phase margin is degraded to around 75º leaving plenty
of margin for unforeseen effects.
Linearity in this case means how well the output current tracks the input voltage. The
easiest way to verify this is by looking at the output voltage versus input voltage since
the voltage over the resistor determines the output current. The results of this can be
seen in Figure 6.21. The converter is linear within 0.5% from 100 mV to 850 mV. The
error around its intended working voltage of 500 mV is 0.02%. The degradation in
performance at lower and higher voltages is caused by lower loop-gain in the opamp
when its output voltage comes closer to the supply.
Temperature versus out current if plotted in Figure 6.22. It shows that the output
current changes around 90 nA from 0-110ºC. This equates to a 0.2 % change in output
current.
46
Figure 6.20 Bode plot of the voltage-to-current converter
47
6.2.6 Voltage-to-current converter bias circuitry
6.2.6.1 Theory
The opamp used in the V-I converter is a folded-cascode architecture. This means that
it needs four different bias voltages. Two for the current source on the N- and P-side
respectively and two for the cascodes on each side.
The bias net used is depicted in Figure 6.23 below.
This type of bias circuit is called stable transconductance [1]. This is because it keeps
gm of the transistors in the opamp stable regardless of supply voltage, process and
temperature variations. It works by stabilizing the transistors transconductance to the
resistors conductance as shown in [1].
6.2.6.2 Implementation
The four rightmost transistors work as a startup circuit. The startup circuit is
necessary because the bias circuitry has two stable states, one being its normal
operating region the other being zero current.
R0 is the resistor that stabilizes the transistor transconductance. In this case it’s
realized with an on-chip resistor. It could be realized with an off-chip resistor but the
external parasitics can cause oscillation. The disadvantage to using an on-chip resistor
is the fact that its value is dependant on temperature and that its absolute value will
have a large variation. The temperature coefficient is fairly low and shouldn’t be a
problem though.
6.2.6.3 Results
The performance aspects of interest can only be tested together with the V-I
converter. If it works the way it’s supposed to it shouldn’t have any effect on the
performance of the V-I converter.
48
7 Results and conclusion
49
Another difficulty during the final stages of simulation is the impact of bond wires on
performance. This is mainly because there are no good models for the bond wires and
a small change in the model leads to large changes in performance.
7.4 Conclusions
Is has been shown that it’s possible to design a 10-bit DAC in a 0.18µm CMOS
process. It has been shown that there are three limits on the update rate of the
converter and one of them will decide the final update rate. The first limit is output
impedance, which places a limit on the highest output frequency that can be
reproduced without too much distortion. This doesn’t really limit the update rate
unless operation up to Nyquist is wanted. In this converter the maximum output
frequency is limited to 635 MHz, which, if operation up to Nyquist is desired, limits
the update frequency to 1.27 GHz. The second limit is imposed by the digital net
because of its delay. The data from the digital net must be available within half a
clock cycle. The delay through the digital net is 695 ps, which limits the update rate to
720 MHz. Another limiting factor is the slew-rate of the output or settling time. The
DAC has to settle before the next step is taken. The settling time is around 2 ns, which
limits the update frequency to 500 MHz. Finally there is actually one more limit on
50
update frequency that is much harder to decide. That is the one caused by noise on
supply lines and ringing on the output both of which are caused by bond wires.
The biggest challenge is to accurately model all errors that are not shown in a circuit
simulator such as noise and mismatch. There are also many other unknown factors
that might degrade performance such as substrate coupling, crosstalk between digital
and analog and the impact of bond wires.
51
52
8 Measurements
8.1 Introduction
All measurements have been performed with a single-ended load of 50 Ω and a full-
scale current of 5 mA. During dynamic measurements the differential signal was
converted to a single-ended signal using a balun. The output signal was analyzed
using a spectrum analyzer. During the static measurements the differential-to-single-
ended conversion was done with a precision opamp and the output was measured with
a multimeter.
53
As can be seen DNL is within ±0.4 LSB and INL within ±0.6. Due to noise during the
measurements the measured results are slightly pessimistic. To make more accurate
measurements it would have been preferable if one could have taken the mean value
of several measurements instead of just one. Due to time constraints this was not
possible, however, a short measurement series was done by taking the mean value of
4 measurements on each level for the first 100 data values. It could be seen that the
performance was slightly better than the one presented above.
0.756 78 77 77
1.7 73 73 73 73
3.3 73 72 71
5.5 67 68 69 67
7.3 65 70 68
9.7 60 69 67 63
13 57 66 65
17 54 62 65 62
22 61 61
37 55 54 58
47 57
61 52 52
86 49
7.3 -62.7
9.7 -60.0
54
Figure 8.5 ENOB versus signal frequency
From Table 8.1 and 8.2 it can be seen that the performance is good for all but the
highest frequencies. Figure 8.5 shows the ENOB versus signal frequency at 80 MSPS.
It can be seen that the performance is good all the way up to 22 MHz. There were
some problems during the measurements with the equipment used. Due to the fact that
the clock from the pattern generator deteriorated the performance an external clock
source was used. The external source triggered the pattern generator, which sent the
data to the DAC. It proved to be difficult getting this to work perfectly due to the
internal delay of the pattern generator. This might have caused some deterioration of
the measured performance.
55
Below some typical output spectrums are presented:
Figure 8.6 Measured output spectrum at 80 MSPS with a 9.7 Mhz input signal
Figure 8.7 Measured output spectrum at 80 MSPS with a 22 Mhz input signal
56
9 References
[1] David A. Johns and Ken Martin, ”Analog Integrated Circuit Design”, John
Wiley & Sons Inc, ISBN 0-471-14448-7.
[2] Mikael Gustavsson, J. Jacob Wikner and Nianxiong Nick Tan, “CMOS Data
Converters for Communications”, Kluwer Academic Publishers, 2000, ISBN 0-
7923-7780-X.
[3] A. Van den Bosch, M. Steyaert & W. Sansen “SFDR-Bandwidth limitations for
high-speed high-resolution current-steering CMOS D/A converters”, in Proc.
IEEE Int. Conf. Electronics, Circuits and Systems (ICECS), Sept. 1999,
pp.1193 –1196.
[4] Anne Van den Bosch, Marc A. F. Borremans, Michel S. J. Steyaert and Willy
Sansen, ”A 10-bit 1-Gsample/s Nyquist Current-Steering CMOS D/A
Converter”, IEEE Journal of Solid-state Circuits Vol. 36, March 2001.
[5] Yasuyuki Nakamura, Takahiro Miki, Atsushi Maeda, Harufusa Kondoh and
Nobuharu Yazawa, ”A 10-b 70-MS/s CMOS D/A Converter”, IEEE Journal of
Solid-state Circuits Vol. 26, April 1991.
[6] Marcel J. M. Pelgrom, AAD C. J. Duinmaijer, Anton P. G. Welbers, ”Matching
Properties of MOS Transistors”, IEEE Journal of Solid-state Circuits Vol. 24,
October 1989.
[7] K. Ola Andersson, Niklas U. Andersson, Mark Vesterbacka, and J. Jacob
Wikner, “A 14-Bit Dual Current-Steering DAC”, SSoCC, 2003.
[8] H. Hassander and C. Lindholm, ”Simulering, layout och verifiering av mixed-
mode-kretsar i Cadence”, Del av rapport i IC-projekt – Electroscience LTH ,
Maj 2003.
[9] Jan M. Rabaey, “Digital Integrated Circuits a Design Perspective”, Prentice
Hall International, 1996, ISBN 0-13-394271-6.
[10] Hironori Banba, Hitoshi Shiga, Akira Umezawa, Takeshi Miyaba, Toru
Tanzawa, Shigeru Atsumi and Koji Sakui, ”A CMOS Bandgap Reference
Circuit with Sub-1-V Operation”, IEEE Journal of Solid-state Circuits Vol. 34,
May 1999.
[11] Johan H. Huijsing, “Operational Amplifiers Theory and Design”, Kluwer
Academic Publishers, 2001, ISBN 0-7923-7248-0.
57
58
Appendix A
The following pages shows the Taylor expansion done in Maple. The first one is the
single-ended case and the other one is the differential.
Single-ended:
//The output signal as a function of gimp=1/Rl
> f:=N*Ir*(x+1)/(2*gl+gimp*N*(x+1));
59
//Set the demand on SFDR to at least 60 db
> Q:=simplify(HDS/A)=.001;
Differential:
//The output signal as a function of gimp=1/Rl
> f:=N*Ir*(x+1)/(2*gl+gimp*N*(x+1));
60
//Replace x with sin(wt) to be able to identify harmonics
> tayt:=subs(x=sin(wt),taydiff);
61
62
Appendix B
Here are the MATLAB programs that were written to simulate the function and the
impact of different errors on the DAC.
function [SFDRar,nobbar,maxinl,maxdnl,Itot]=
DAC(nSample,Fsamp,Fsig,overtoner,noit,funk,sv);
%ADC
%Ramp
if (funk==1)
data = [0:1:1023];
end
%Sin
if (funk==2)
w = 2*pi*Fsig;
tid = [0:nSample-1]./Fsamp;
brus = 0.3*rand(size(tid));
data = round(1023*(1+sin(w*tid))/2+brus);
end
%OFDM
if (funk==3)
Fsig=(312.5e3:312.5e3:26*312.5e3);
w=2*pi.*Fsig;
tid=[0:nSample-1]./Fsamp;
phi=2*pi*rand(26);
sig=zeros(26,nSample);
for i=1:26
sig(i,:)=0.07807*sin(w(i)*tid+phi(i));
end
sigtot=sum(sig);
brus = 0.3*rand(size(tid));
data = round(1023*(1+sigtot)/2+brus);
end
%DAC
Iref=5e-3/N;
sd=0.502/100;
gradient=4/400;
gradpolx=5/400;
gradpoly=5/400;
Lnoise=6.3e-6;
Inoise=3.03e-7;
%i=1;
%for k=1:16
% for l=1:16
% imatris(k,l)=i;
63
% i =i+1;
% end
%end
%Generera gradienter
t=[-1:2/31:1];
xpolgrad=-gradpolx*t.^2-gradpolx;
ypolgrad=-gradpoly*t.^2+gradpoly;
xgrad=[0.5-gradient:2*gradient/31:0.5+gradient]+xpolgrad;
ygrad=[0.5-gradient:2*gradient/31:0.5+gradient]+ypolgrad;
for k=1:32
for l=1:32
gradmat(k,l)=xgrad(k)+ygrad(l);
end
end
%Starta iterationsprocessen
for j=1:noit
%Generera strömkällor
Irefs=Iref+Iref*sd*randn(32,32);
Ibin=Iref+Iref*sd*rand(1,3);
Ibinp=Ibin;
Ibinn=fliplr(Ibinp);
Itot=Irefs.*gradmat;
%Spara strömmatrisen
if (sv==1)
save strom_matris Itot
else
load strom_matris
end
%Generera brus
Noisetot=sqrt(N*Inoise^2+2*Lnoise^2)*randn(size(data));
%Räkna fram total ström genom att räkna fram varje summa av
strömkällor
Iutsump=zeros(1,1024);
Iutsumn=zeros(1,1024);
for i=1:4:1020
for k=1:3
Iutsump(i+k)=Iutsump(i+k-1)+Ibinp(k);
Iutsumn(i+k)=Iutsumn(i+k-1)+Ibinn(k);
end
64
rowp=imsp(fix(i/4)+1,1);
colp=imsp(fix(i/4)+1,2);
rown=imsn(fix(i/4)+1,1);
coln=imsn(fix(i/4)+1,2);
%Summera källorna enligt common-centroid (varje kvadrant flippad
%så att enskilda källor får gemensamt centrum)
Iutsump(i+4)=Iutsump(i)+Itot(rowp,colp)+Itot(rowp,33-colp)+Itot(33-
rowp,33-colp)+Itot(33-rowp,colp);
Iutsumn(i+4)=Iutsumn(i)+Itot(rown,coln)+Itot(rown,33-coln)+Itot(33-
rown,33-coln)+Itot(33-rown,coln);
%Summera källorna utlagda likadant (ingen rotation av kvadranter)
%Iutsump(i+4)=Iutsump(i)+Itot(rowp,colp)+Itot(rowp,17-
colp)+Itot(17-rowp,17-colp)+Itot(17-rowp,colp);
%Iutsumn(i+4)=Iutsumn(i)+Itot(rown,coln)+Itot(rown,17-
coln)+Itot(17-rown,17-coln)+Itot(17-rown,coln);
end
for k=1022:1024
Iutsump(k)=Iutsump(k-1)+Ibinp(k-1021);
Iutsumn(k)=Iutsumn(k-1)+Ibinn(k-1021);
end
for i=1:length(data)
Ioutp(i)=Iutsump(data(i)+1);
Ioutn(i)=Iutsumn(1024-data(i));
end
%Räkna ut utspänningen
voutp= (2*Ioutp)./((2/Rl)+(1/Rimp)*2*data);
voutn= (2*Ioutn)./((2/Rl)+(1/Rimp)*(2046-2*data));
Idiff=Ioutp-Ioutn;
vdiff=voutp-voutn+Noisetot;
Vref1=max(vdiff)-min(vdiff);
Vref=600e-3;
inl=(vdiff'-A*c)./(Iref*Rl);
for m=1:length(data)-1
dnl(m)=(vdiff(m+1)-vdiff(m)-2*Iref*60)/(Iref*60);
end
maxdnl(j)=max(abs(dnl));
MaxDNL=maxdnl(j)
maxinl(j)=max(abs(inl));
MaxINL=maxinl(j)
SFDRar=0;
nobbar=0;
j
end
matrX=[1:32];
matrY=[1:32];
if funk==2
fullfft;
65
FOM
nob1=FOM(63:67);
SFDRvar=FOM(44:48)
figure(1)
SFDRvar=str2num(SFDRvar);
SFDRar(j)=SFDRvar;
nob1=str2num(nob1);
nobbar(j)=nob1;
maxinl=0;
maxdnl=0;
j
end
if funk==3
fullfft;
FOM
nobbar=0;
SFDRar=0;
maxinl=0;
maxdnl=0;
end
end
66
function m=swmatris();
m=zeros(4,8);
mtmp=zeros(8,8);
i=1;
for j=1:4
m(i,j)=256 - 4*(j-1);
end
for k=1:4
m(i,j+k)=m(i,j) + 4*(k-1)-1;
end
for i=2:4
for j=1:8
m(i,j)=m(1,j)-64*(i-1);
end
end
m2=flipud(m) - 16;
mtmp=[m;m2];
mtmp4=fliplr(flipud(mtmp));
mtmp1=fliplr(flipud(mtmp4)) - 2;
mtmp2=fliplr(mtmp1) - 30;
mtmp3=fliplr(flipud(mtmp2)) - 2;
67
clear;
nSample=4096;
Fsamp=80e6;
Fsig=11e6;
overtoner=9;
noit=100;
Vref=600e-3;
funk=2;%1=Ramp 2=Sinus 3=OFDM
save_yes=1;%1=Skapa och spara stromkallor
[SFDR,ENOB,maxinl,maxdnl,Itot]=DAC(nSample,Fsamp,Fsig,overtoner,noit,
funk,save_yes);
matrX=[1:32];
matrY=[1:32];
%figure(1)
%surf(matrX,matrY,Itot);
matrX=[1:noit];
matrY=[1:noit];
if funk==1
figure(2)
hist(maxinl,50);
xlabel('INL [LSB]');
ylabel('Number of DACs')
figure(3)
hist(maxdnl,50);
xlabel('DNL [LSB]');
ylabel('Number of DACs')
end
if funk==2
figure(2)
hist(SFDR,50);
xlabel('SFDR [dB]');
ylabel('Number of DACs')
figure(3)
hist(ENOB,50);
xlabel('ENOB [Number of bits]')
ylabel('Number of DACs')
end
68
Appendix C
for( k 1 nbrElements
dbCreateInst(cvSch NAND nil list(x (-nbrElements - k))
"R0" 1)
)
else
for( k 1 nbrElements
dbCreateInst(cvSch NAND nil list(x (1-k)) "R0" 1)
)
for( k 1 nbrElements
dbCreateInst(cvSch NOR nil list(x (-nbrElements - k))
"R0" 1)
)
) ; ** if mod **
69
if( (k == (expt(2 (i - 1)) - 1))
then
x1 = x1 - (NYx - IYx)
y1 = y1 - (NYy - IYy)
) ; ** if k **
) ; ** for k **
) ; ** for k **
x1 = x - 0.25
x2 = x1
y1 = NB
y2 = (-nbrE + 1) + NB
schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2) 0.0625
0.0625 0.0)
for( k 0 (nbrElements - 1)
x1 = old_x + NYx
x2 = x
y1 = -k + NYy
y2 = -(nbrElements + 1) + NA - k
if( (k == ((nbrElements - 1) / 2))
then
x1 = old_x + IYx
y1 = -k + IYy
) ; ** if k **
) ; ** for k **
x1 = x
x2 = -2
y1 = -((nbrE - 1) / 2) + IA
y2 = y1
id = schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2)
0.0625 0.0625 0.0)
id = car(id)
sprintf(data_in "data_in<%d>" i)
xL = x2
yL = y1 + 0.0625
70
schCreateWireLabel(cvSch id (xL:yL) data_in "lowerLeft" "R0"
"fixed" 0.1 nil)
) ; ** if i **
if( (i == nbr )
then
nbrE = expt(2 (i + 1)) - 1
x2 = x + NYx + 1
for( k 0 (nbrE - 1)
x1 = x + NYx
y1 = -k + NYy
y2 = y1
if( (k == ((nbrE - 1) / 2))
then
x1 = x + IYx
y1 = -k + IYy
y2 = y1
) ; ** if k **
) ; ** for k **
) ; ** if i **
if( (i == 1)
then
x1 = 0
x2 = -0.25
for( k 0 2
y1 = -k + NB
y2 = y1
if( (k == 1)
then
y1 = -k + IA
y2 = y1
) ; ** if k **
schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2)
0.0625 0.0625 0.0)
) ; ** for k **
x1 = x2
y1 = NB
y2 = NB - 2
schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2) 0.0625
0.0625 0.0)
x1 = 0
x2 = -2
y1 = NA
y2 = y1
71
id = schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2)
0.0625 0.0625 0.0)
id = car(id)
xL = x2
yL = y1 + 0.125
schCreateWireLabel(cvSch id (xL:yL) "data_in<0>" "lowerLeft"
"R0" "fixed" 0.1 nil)
y1 = IA - 1
y2 = y1
yL = y1 + 0.125
id = schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2)
0.0625 0.0625 0.0)
id = car(id)
schCreateWireLabel(cvSch id (xL:yL) "data_in<1>" "lowerLeft"
"R0" "fixed" 0.1 nil)
x1 = -0.5
x2 = 0
y1 = NA
y2 = NA - 2
schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2) 0.0625
0.0625 0.0)
) ; ** if i **
old_x = x
x = expt(i 3) + 4
) ; ** for i **
x1 = -4
x2 = x1 - 2
id = schCreateWire(cvSch "route" "full" list(x1:y1 x2:y2) 0.0625
0.0625 0.0625)
id = car(id)
msb = nbr
sprintf(data_in "data_in<%d:0>" msb)
xL = x2
yL = y1 + 0.125
schCreateWireLabel(cvSch id (xL:yL) data_in "lowerLeft" "R0"
"fixed" 0.15 nil)
) ; ** procedure bin2term **
72
Linköping Studies in Science and Technology
Thesis No. 976
STUDIES ON PERFORMANCE
LIMITATIONS IN CMOS DACS
K Ola Andersson
LiU-Tek-Lic-2002:49
i
CMOS is also the preferred technology for implementing large systems on a sin-
gle chip with both analog and digital blocks. Three different current-steering
CMOS DACs were developed in this work, and are presented in the thesis. Mea-
surement results show close resemblance with the simulation results obtained
from the developed models.
ii
Acknowledgments
Many people deserve my gratitude for helping me during the years. First of all I
thank my supervisor, Prof. Mark Vesterbacka, for his guidance.
Dr. J Jacob Wikner, has taught me a lot about data conversion, mixed signal
design, IT tools, and other things that have simplified my work. All help from
him, M.Sc. Niklas U Andersson, and Dr. Mikael Karlsson Rudberg, all from Infi-
neon Technologies Wireless Solutions Sweden (former Ericsson Microelectron-
ics), is greatly appreciated.
I also thank Dr. Gunnar Björklund and M.Sc. Magnus Hägglund for supporting
my work during the years I spent doing research at Ericsson Microelectronics.
I appreciate the help and inspiration from all my colleagues at Ericsson Micro-
electronics and Electronics Systems, Linköping University.
Finally, I thank my family, especially my wife Helena, for their support.
The work was financially supported by the Microelectronics Research Center
(MERC) at Ericsson Microelectronics and the Center for Industrial Information
Technology (CENIIT) at Linköping University.
iii
iv
Table of Contents
Chapter 1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The analog front end . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Scientific contribution of the author . . . . . . . . . . . . . . . . . . . . . 3
1.4 Publications related to the author . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.1 Conference publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4.2 Journal publication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.3 Patents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.4 Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Chapter 2
DACs in Telecommunication . . . . . . . . . . . . . . . . . . . . 7
2.1 The DSL Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Signals and modulation in ADSL . . . . . . . . . . . . . . . . . . . . 9
Distorted signals 10
2.1.2 DAC requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 The ideal DAC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Representation of digital signals . . . . . . . . . . . . . . . . . . . 12
2.2.2 Digital-to-analog conversion . . . . . . . . . . . . . . . . . . . . . . 13
2.2.3 Nyquist rate DACs vs. oversampled DACs . . . . . . . . . . . 14
v
2.3 DAC performance measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Performance measures in the code domain . . . . . . . . . . . 16
Differential nonlinearity, DNL 16
Integral nonlinearity, INL 17
2.3.2 Measures in the frequency domain . . . . . . . . . . . . . . . . . . 17
Signal-to-noise ratio, SNR 19
Signal-to-noise-and-distortion ratio, SNDR 19
Effective number of bits, ENOB 19
Spurious-free dynamic range, SFDR 20
Total harmonic distortion, THD 20
Multi-tone power ratio, MTPR 20
Chapter 3
Modeling of Current-Steering DACs . . . . . . . . . . . . . 23
3.1 Classification of systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 The ideal current-steering DAC . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Modeling of static errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3.1 Component mismatch . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Influence of graded element matching errors 28
Modeling of random matching errors 34
3.3.2 Finite output resistance . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4 Modeling of dynamic errors . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4.1 State-space model of a current-steering DAC . . . . . . . . . 35
Modeling of nonideal DAC components 35
Circuit-level model of the DAC 38
Simulations 42
3.4.2 A low-complexity model . . . . . . . . . . . . . . . . . . . . . . . . . 44
A simple DAC model 45
Model structure 46
Simulation of the low-complexity model 48
3.5 Combined models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Chapter 4
Compensation and Correction of Errors . . . . . . . . . . 51
4.1 Compensation and correction of static errors . . . . . . . . . . . . 51
4.1.1 Calibration of the MSBs . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Basic idea 52
vi
Proposed implementation 54
Qualitative comparison with other calibration techniques 55
4.1.2 Distributed biasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.1.3 Dynamic element matching . . . . . . . . . . . . . . . . . . . . . . . 56
Generalized DEM 57
DEM utilizing switching trees 58
Partial randomization DEM 59
Mismatch shaping DEM 60
4.2 Compensation and correction of dynamic errors . . . . . . . . . . 61
4.2.1 Differential DACs with variable common mode . . . . . . . 62
Proposed redundant architecture 62
Dithering of the common-mode level 64
DC level minimization 64
Common-mode level reduction with boundary conditions 66
4.2.2 Modulation of expected errors . . . . . . . . . . . . . . . . . . . . . 67
Delta-sigma modulator basics 68
Spectral shaping of nonlinearities 70
Simulations 71
Measurements 72
4.2.3 Predistortion of dynamic errors . . . . . . . . . . . . . . . . . . . . 72
Predistortion block 73
Simulation results 74
4.2.4 Implementation issues . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Chapter 5
Current-Steering DAC Implementations . . . . . . . . . . 77
5.1 CMOS processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.1.1 Large signal models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
NMOS transistors 78
PMOS transistors 79
Notes on large signal models 80
5.1.2 Small signal models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2 Analog DAC building blocks . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2.1 CMOS current sources . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Single transistor current source 81
Cascode current source 82
5.2.2 CMOS switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Digital DAC building blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3.1 Binary-to-thermometer encoding . . . . . . . . . . . . . . . . . . . 84
vii
Thermometer coded and segmented structures in general 84
Implementation of binary-to-thermometer encoders 87
5.3.2 Switch signal generators . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 DAC design strategies and measurement setup . . . . . . . . . . . 91
5.4.1 Overall layout structure . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.4.2 Clocking strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4.3 Measurement setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.5 A 14-bit DAC in 0.35 µm CMOS . . . . . . . . . . . . . . . . . . . . . . . 94
5.5.1 Chip description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.5.2 Measurement results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.6 A 14-bit PRDEM DAC in 0.35 µm CMOS . . . . . . . . . . . . . . . 97
5.6.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.6.2 Simulations and comparison with measurements . . . . . . 100
Simulation setup and results 100
Measurement results 101
Comparison of SFDR performance 101
5.7 A dual 14-bit DAC in 0.25 µm CMOS . . . . . . . . . . . . . . . . . . 102
5.7.1 Architecture and implementation . . . . . . . . . . . . . . . . . . 103
5.7.2 Measurement results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.8 Summary of implemented DACs . . . . . . . . . . . . . . . . . . . . . . 105
Chapter 6
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
viii
1 Introduction
Strict requirements are imposed on the transmitting and receiving circuits for
broadband communication. A bottle neck in these communication systems is the
data converters, i.e., digital-to-analog converters (DACs) and analog-to-digital
converters (ADCs). In this work we study performance limitations in DACs for
high-speed applications. This chapter is an overview of the work and an introduc-
tion to the thesis. In Sec. 1.1 we give a brief introduction to the analog front end
(AFE) in which the DAC is an important building block. The different parts of
the thesis are outlined in Sec. 1.2. The author’s scientific contribution to the area
is outlined in Sec. 1.3, and the publications related to this thesis are listed in
Sec. 1.4. This chapter ends in Sec. 1.5 with a list of abbreviations used through-
out the thesis.
1
Introduction
ing the analog data to digital. The digital data is processed by a fast Fourier trans-
form (FFT) circuit to determine the correct points in the different constellations,
as will be discussed in Sec. 2.1.1.
DAC LP A
DSP blocks channel
ADC LP A
receive amplifier
2
Scientific contribution of the author
3
Introduction
[3] K.O. Andersson, N.U. Andersson, and J.J. Wikner, “Spectral shaping of
DAC nonlinearity errors through modulation of expected errors,” Proc.
IEEE International Symposium on Circuits and Systems (ISCAS’01),
vol. 3, pp. 417-420, Sydney, Australia, May 6-9, 2001.
[4] N.U. Andersson, K.O. Andersson, J.J. Wikner, and M. Vesterbacka,
“Models and implementation of a dynamic element matching DAC,” Proc.
NORCHIP Conference, pp. 155-160, Kista, Sweden, Nov. 12-13, 2001.
[5] K.O. Andersson, N.U. Andersson, M. Vesterbacka, and J.J. Wikner, “A
differential DAC architecture with variable common-mode level,” Proc.
IEEE International Symposium on Circuits and Systems (ISCAS’02),
vol. 1, pp. 113-116,Scottsdale, AZ, USA, May 26-29, 2002.
[6] K.O. Andersson, N.U. Andersson, M. Vesterbacka, and J.J. Wikner,
“Combining DACs for improved performance,” Proc. 4th IEE
International Conference on Advanced A/D and D/A Conversion
Techniques and their Applications, Prague, Czech Republic, June 26-28,
2002.
[7] M. Vesterbacka, K.O. Andersson, N.U. Andersson, and J.J. Wikner,
“Using different weights in DACs,” Proc. 4th IEE International
Conference on Advanced A/D and D/A Conversion Techniques and their
Applications, Prague, Czech Republic, June 26-28, 2002.
1.4.3 Patents
[9] K.O. Andersson, “Improved current-steering D/A conversion,” Swedish
patent 0000731-0, 2000.
[10] K.O. Andersson, “Current Steering DAC”, pending Swedish patent, 2001.
[11] K.O. Andersson and J.J. Wikner, “Digital-to-analog converter having error
correction,” pending Swedish patent, 2002.
4
Abbreviations
1.4.4 Tutorial
[12] J.J. Wikner, N.U. Andersson, K.O. Andersson, and M. Vesterbacka,
Compilation of error prevention and correction techniques for DACs,
Tutorial IEEE 7th Int. Conf. on Electronics, Circuits, and System
(ICECS’01), Malta, Sep. 2, 2001.
1.5 Abbreviations
Below is a list of the abbreviations used in the thesis.
ADC Analog-to-digital converter
ADSL Asymmetric digital subscriber line
AFE Analog front-end
CMOS Complementary metal oxide semiconductor
CO Central office
CPE Customer premises equipment
DAC Digital-to-analog converter
DCVS Differential cascode voltage switch
DEM Dynamic element matching
DMT Discrete multi-tone
DNL Differential nonlinearity
DSL Digital subscriber line
DSP Digital signal processing
ECH Echo canceled hybrid
ENOB Effective number of bits
FDM Frequency division multiplexing
FFT Fast Fourier transform
FIR Finite impulse response
FRDEM Full randomization dynamic element matching
GPIB General purpose interface bus
IFFT Inverse fast Fourier transform
INL Integral nonlinearity
LP Lowpass
5
Introduction
6
2 DACs in
Telecommunication
The main emphasis in this thesis is on the linearity of current-steering DACs. The
work is focused on converters for telecommunication applications, and this chap-
ter is an introduction to DACs in general, and their application to telecommunica-
tion. We start by outlining the set of signals the DACs are intended for, and
illustrate the importance of high linearity in the intended applications. As appli-
cation examples we use digital subscriber line (DSL) environments. The ideal
DAC is presented on a behavioral level, along with some important performance
measures used to describe the degree of deviation from the ideal DAC.
7
DACs in Telecommunication
since the attenuation of signals in copper wires is rather high for the frequency
bands used. The maximum physical distance between the CPE and the CO side
for ADSL is approximately 5.5 km. Over this limit the possible data rate is dra-
matically reduced. This is due to the common use of so called loading coils on
such long lines to improve the voice band properties of the line. As a side effect
these loading coils increase the attenuation for frequencies above the voice band
[1].
Backbone
CPE CPE
CO CO
CPE
CPE
CPE
In Fig. 2.2 we show the frequency bands used in ADSL, which is a standard that
has been well established and is widespread today. The range of frequencies
below 4 kHz is reserved for POTS. The frequency range between 30 and 138 kHz
is used for upstream communication, i.e., data sent from the CPE side to the CO
side. Data is sent in the reverse direction, downstream, in the frequency range
between 138 kHz and 1.104 MHz. A guard band is present between the POTS
band and the ADSL downstream band. The information in ADSL is transmitted
using a discrete multi-tone (DMT) signal. The ADSL version described here is
known as frequency division multiplexing (FDM) ADSL [1, 2, 3, 4], where the
upstream and downstream bands are separated. There is also a version called
echo-canceled hybrid (ECH) ADSL, where the upstream and downstream bands
overlap. ECH ADSL requires echo cancellation to separate transmitted data from
received data [1, 2, 3, 4].
The downstream capacity (up to 9 Mb/s) is larger than the upstream capacity (up
to 1 Mb/s) since the desired rate for downloading data in some applications, e.g.,
video-on-demand, is larger than the corresponding uploading rate [4], hence the
term asymmetric. In VDSL on the other hand, the communication is symmetric,
i.e., the same data rate is available in both directions.
8
The DSL Environment
The maximum signal bandwidth in ADSL is about 1 MHz. In other types of tele-
communication and data communication, e.g., VDSL and wireless local area net-
work (WLAN), the corresponding bandwidths lie around 10 MHz, and in future
communication standards the bandwidths may be even higher.
ADSL downstream
ADSL upstream
guard band
POTS
0 4 30 138 1104
Frequency [kHz]
It should be noted that, with the notation used in (2.1), some of the tones must
have zero amplitude in ADSL, since their frequencies are in the POTS band or
the guard band (or in the wrong ADSL band for FDM ADSL).
To each carrier there is an associated quadrature amplitude modulation (QAM)
constellation [3] given by the values of a i and b i . The bits that are to be transmit-
ted are divided into different constellations. The number of bits modulated on
9
DACs in Telecommunication
each carrier is not fixed, but is determined during training sequences and depends
on the quality of the transmission link for the corresponding frequency band. In
Fig. 2.3(a) a 16-QAM constellation is shown, and the point of the constellation to
be transmitted determines the actual values of a i and b i .
Distorted signals
When the signal is sent from the transmitter to the receiver it will be distorted.
Sources of signal distortion are, e.g., attenuation in the twisted-pair wire and non-
linearities in the transmitting circuits. Here we examine the impact on the QAM
constellation for different types of distortion using simulations in Matlab. The
simulations are performed using discrete time signals. Seven tones with angular
frequencies Ω n = n ⋅ π ⁄ 8 for n = 1, 2, …, 7 have been used, and to each of
these tones there is a 16-QAM constellation as illustrated in Fig. 2.3(a). As an
example we investigate how the constellation for the tone with n = 3 is affected
by distortion. For the remaining tones we choose random values of a i and b i
with rectangular distribution, and for each such random combination we consider
all possible combinations of a 3 and b 3 . The constellations plotted in Fig. 2.3 are
the results of 100 stochastic outcomes.
First we examine how the constellations are affected by a linear distortion. Linear
distortion affects each carrier by a signal independent phase shift and a signal
independent attenuation (or possibly amplification). Hence, the constellation of
the linearly distorted carrier is a rotated and scaled copy of the original constella-
tion, as illustrated in Fig. 2.3(b). The linear distorter is a simple finite impulse
response (FIR) filter given by y(n) = 0.5 ⋅ x(n) + 0.1 ⋅ x(n – 1) . Since there is a
one-to-one mapping from the original constellation to the distorted constellation,
the original constellation can be restored once the attenuation and the phase shift
is known.
The constellations for signals that are distorted by a simple nonlinearity,
y(n) = x(n) + 0.2 ⋅ x 3(n) , are shown in Fig. 2.3(c). It can be seen that there is no
longer a one-to-one mapping between the original constellation and the distorted
constellation. The location of the received point is depending on the amplitudes
and phases of the other tones as well. When the nonlinearity becomes more
severe, the clouds of points start to overlap and errors occur in determining the
original location of the transmitted point.
With these simple examples we have illustrated the importance of high linearity
in all parts of the transmission link. Noise added to the signal will have similar
effects on the constellations as a nonlinearity, so there are also requirements on
the noise margins.
10
The DSL Environment
Original constellation
1.5
0.5
b
−0.5
−1.5
(a)
Linearly distorted constellation Nonlinearly distorted constellation
1.5 1.5
0.5 0.5
b
−0.5 −0.5
−1.5 −1.5
(b) (c)
Figure 2.3 (a) 16-QAM constellation, (b) linearly distorted constellation, and (c) nonlin-
early distorted constellation.
General
Property
requirement
Signal bandwidth 1-10 MHz
Effective number of bits 12-14
Table 2.1 General DAC requirements for DSL and similar applications.
11
DACs in Telecommunication
A comment on the requirements is that 12-bit DACs are commonly used in FDM
ADSL, whereas ECH ADSL often requires 14 bits of accuracy [1]. For some
emerging communication standards the bandwidths are much higher than in
Table 2.1 [5].
since they easily can be distinguished by the context. The binary valued signals
b k(n) are the individual bits of the digital word X (n) , and w k is the weight asso-
ciated with bit k . The weights depend on what coding scheme is used. For exam-
ple, in the binary-coded case we have that w k = 2 k , whereas w k = 1 for all k in
the thermometer-coded case. These two types of codes are common in data con-
version, and are more thoroughly described in Sec. 5.3.1. We use the notation b
for the inverse of b , and X denotes X with all individual bits inverted.
Throughout this thesis we represent the weights, w k , as positive numbers. That
is, we treat digital numbers and signals as positive. In the DSP blocks, digital sig-
nals are generally expressed as signed numbers, using, e.g., two’s complement
representation [6]. However, all designs of DACs in this thesis operate by sum-
ming positive analog weights, so from a DAC designer’s point of view it is more
convenient to use positive weights also in the digital domain.
12
The ideal DAC
1 for 0 ≤ t < T
P(t) = . (2.7)
0 otherwise
When designing a DAC, the goal is to come as close as possible to the operation
described in (2.6) and (2.8). It is, however, not possible to achieve exactly the
operation of an ideal DAC, since A(t) given by (2.6) is discontinuous (unless
X (n) is constant). The sinc pulse and the unit square pulse are plotted together in
Fig. 2.4.
13
DACs in Telecommunication
1
Amplitude level
−5 −4 −3 −2 −1 0 1 2 3 4 5
Normalized time, t/T
Figure 2.4 The sinc pulse (dashed) and the unit square pulse (solid) plotted together.
fu = 2 ⋅ f0 (2.9)
as touched upon in the previous section. A DAC using this update frequency is
referred to as a Nyquist rate DAC.
In an ideal data converter, the only errors are those caused by the quantization
process, referred to as quantization noise. It is sometimes argued that there is no
quantization noise in a DAC, since the input signal is already quantized [8]. How-
ever, since the input is already quantized it contains quantization noise. We can
therefore talk about quantization noise in the output signal from a DAC, even if
the quantization is a process in the circuits preceding the DAC.
For an N -bit DAC with N relatively large and with a varying input signal, the
quantization noise, q(n) , can often be adequately described as a white noise sig-
nal with rectangular distribution between 0 and 1 (assuming that the size of the
quantization step is unity). The total power of the quantization noise is then [7, 9]
14
The ideal DAC
1
P q = Var(q(n)) = ------ . (2.10)
12
This power is evenly distributed over frequencies in the interval ( 0, f u ⁄ 2 ) (since
the noise is treated as white). If we use an update frequency larger than the
required 2 ⋅ f 0 , some of the noise will appear outside of the signal band, at fre-
quencies in the interval ( f 0, f u ⁄ 2 ) . This part of the noise can be filtered out,
since the signal does not have any spectral content for these frequencies, and the
overall signal-to-noise-and-distortion ratio (SNDR) (see Sec. 2.3.2) is improved.
A DAC with f u > 2 ⋅ f 0 is known as an oversampled DAC, and the oversampling
ratio (OSR) is defined as
fu
OSR = ------------
-. (2.11)
2 ⋅ f0
Compared with a Nyquist rate DAC, the signal-to-noise ratio (SNR) due to quan-
tization noise is increased with 10 ⋅ log 10(OSR) dB [8]. To further decrease the
power of the quantization noise within the signal band, delta-sigma modulation
techniques are often used. These techniques utilize feedback loops to spectrally
shape the quantization noise to frequencies outside of the signal band, and are
briefly outlined in Sec. 4.2.2.
Another reason for using oversampling is that using square pulses instead of sinc
pulses for reconstruction affects the spectral properties of the output signal. Let
A sinc(ω) denote the power spectrum of the signal reconstructed with sinc pulses,
and let A square(ω) denote the corresponding spectrum for the signal reconstructed
using square pulses. Then, the following equality holds [7, 8]
ω
A square(ω) = A sinc(ω) ⋅ sinc(------------) , (2.12)
2π f u
i.e., the signal is attenuated for frequencies close to f u . For the Nyquist fre-
quency, f u ⁄ 2 , this attenuation is ( sinc(1 ⁄ 2) ) –1 = π ⁄ 2 ≈ 3.9 dB . For a Nyquist
rate DAC this attenuation may need to be compensated for with so called anti-
sinc filters, depending on the requirements of the application. For oversampled
DACs there is no spectral content of the signal close to the Nyquist frequency, if
high enough OSR is used. Therefore, the attenuation becomes neglectable.
Moreover, the images of the signal appearing around multiples of the update fre-
quency need to be filtered out with image rejection filters before the signal enters
the communication channel to minimize the interference with other types of
communication utilizing other frequency bands. In an oversampled DAC the
15
DACs in Telecommunication
images are better separated from the signal, compared with Nyquist rate DACs,
and the requirements on the transition band widths of the image rejection filters
are reduced.
15
14
13
12
11
10
Output level
9
8
7
6
5
4
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Input code
16
DAC performance measures
A(X ) – A(X – 1) – K
DNL(X ) = ------------------------------------------------- (2.13)
K
The DNL for the DAC with the transfer characteristics in Fig. 2.5 is plotted in
Fig. 2.6.
1
DNL
−1
−2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Input code
A(X ) – Y (X )
INL(X ) = ----------------------------- . (2.14)
K
Combining (2.13) and (2.14) yields that
17
DACs in Telecommunication
INL 1
−1
−2
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Input code
monic distortion in the output signal. The second and third harmonics are indi-
cated in Fig. 2.8. Hence, the different measures of distortion presented in this
section is also measures of the DAC linearity.
fundamental
2:nd harmonic
Frequency
Figure 2.8 PSD plot of the output signal from a nonideal DAC with a single-tone input sig-
nal.
The signal impurities can be divided into noise and distortion. Noise is typically
signal independent, due to, e.g., thermal activity in circuit elements, having a
smooth spectral appearance. Distortion, on the other hand, is generally character-
18
DAC performance measures
ized by clearly visible peaks in the output spectrum and is often signal dependent,
e.g., due to circuit nonlinearities. Gray zones are present in the analysis, e.g.,
quantization errors are clearly signal dependent, but are often treated as sources
of noise. The total output power is
where P signal is the power of the wanted output signal, P noise is the noise power,
and P dist is the total power due to distortion.
P signal
-.
SNR = -------------- (2.17)
P noise
P signal
SNDR = ------------------------------
-. (2.18)
P noise + P dist
This is a more adequate measure of signal quality than SNR if a large amount of
distortion is present. The SNDR due to quantization noise of the ideal N -bit
(Nyquist rate) DAC with a full-scale sinusoidal input signal can be expressed as
[7, 8]
SNDR – 1.76
ENOB = -------------------------------- , (2.20)
6.02
which is only a different way of expressing the SNDR. It is a useful measure to,
e.g., find a lower bound on the number of bits, N , to achieve a given SNDR.
19
DACs in Telecommunication
P signal
-,
SFDR = -------------- (2.21)
P ls
where P ls is the power of the largest spurious (unwanted) tone. SFDR is indi-
cated in Fig. 2.8. It should be mentioned that this definition of SFDR, which is
often used for data converters, is different from the one most often used for
amplifiers, in which case SFDR is defined as the SNR for the maximum output
power without having any spurious tones [8].
∑ Pn
n=2
THD = ---------------
-, (2.22)
P signal
i.e., THD is the power of all harmonics, normalized with respect to the signal
power.
20
DAC performance measures
MTPR
PSD
Frequency
Figure 2.9 PSD plot of a multi-tone signal with one tone left out.
21
DACs in Telecommunication
22
3 Modeling of Current-
Steering DACs
In this chapter we discuss modeling of current-steering DACs. Both static errors,
such as component mismatch, and dynamic errors, e.g., due to parasitic capaci-
tance are discussed. The study is focused on the dynamic errors.
Modeling of nonlinear behavior in analog circuits is often made using polyno-
mial models, such as Taylor or Volterra series expansions [12, 13]. Some initial
attempts of characterizing DACs using Volterra models were made in this work.
These were, however, unsuccessful in capturing the nonlinear behavior of the
current-steering DAC, except for in very limited frequency bands. Hence, the
polynomial type models have been discarded in this work.
• Suppose that we have two arbitrary input signals X 1(n) and X 2(n) , yielding
the output signals Y 1(n) and Y 2(n) , respectively. Then the system is called
linear if the input signal X (n) = aX 1(n) + bX 2(n) yields the output signal
Y (n) = aY 1(n) + bY 2(n) , where a and b are arbitrary constants. Otherwise,
the system is nonlinear.
23
Modeling of Current-Steering DACs
• Consider the special case for the signals above when X 2(n) = X 1(n – k) ,
where k is an arbitrary integer. The system is time-invariant if
Y 2(n) = Y 1(n – k) , otherwise the system is time-variant.
• The system is called causal if the value of the output signal, Y (n 1) , at instant
n 1 only is a function of the input signal X (n) for n = n 1, n 1 – 1, n 1 – 2, …
(i.e., future values of X (n) have no influence on the output).
• The system is called static if the value of the output signal, Y (n 1) , at instant
n 1 only is a function of the input signal X (n) for n = n 1 . Otherwise, the sys-
tem is called dynamic.
I i = w i ⋅ I unit , (3.1)
where I unit is called the unit current. The DAC shown in Fig. 3.1 utilizes a differ-
ential architecture. The switches are controlled by the digital input such that the
current from source I i is directed to the analog output terminal A + when the dig-
ital input bit b i = 1 , otherwise the current is directed to terminal A – . The total
positive output current is
M–1 M–1
I+ = ∑ bi ⋅ I i = I unit ⋅ ∑ bi ⋅ wi = I unit ⋅ X . (3.2)
i=0 i=0
24
The ideal current-steering DAC
I0 Ij IN1
b0 bj bN1
I+ I
A+ A
RL RL
If we let the weights, w i , be the same as for the digital bits, b i , and compare
expression (3.2) with expression (2.8) in Chapter 2, we find that the output cur-
rent I + can be regarded as the output from an ideal DAC. Similarly, the output
current I – can be regarded as the output from an ideal DAC with input
X = [ b N – 1, …, b 0 ] .
Differential architectures are often used to reject noise and distortion [8]. The dif-
ferential output current, I diff , is defined as
M–1
I diff = I + – I – = 2 ⋅ I unit ⋅ ∑ bi ⋅ wi – I unit ⋅ ∑ wi =
i=0
= ( 2 ⋅ I unit ⋅ X – I unit ⋅ X max ) , (3.4)
where X max is the value of X when all bits b i = 1 . From (3.4) we see that the
differential output current can be regarded as the output from an ideal DAC. We
also refer to the outputs in terms of voltages, i.e.,
V + = R L ⋅ I + = R L ⋅ I unit ⋅ X , (3.5)
V – = R L ⋅ I – = R L ⋅ I unit ⋅ X , (3.6)
and
25
Modeling of Current-Steering DACs
µ ⋅ C ox W 2
I D = ---------------- ⋅ ----- ⋅ ( V gs – V T ) . (3.8)
2 L
The parameter µ is the charge carrier (electron) mobility which is dependent on,
e.g., doping concentration and temperature [14]. C ox is the oxide capacitance per
unit area, and depends on, e.g., the oxide thickness. W and L are the width and
the length of the transistor channel, V gs is the gate-source voltage, and V T is the
threshold voltage, which is dependent on, e.g., the bulk-source voltage of the
transistor. A closer description of MOS transistors is given in Sec. 5.1.
If these transistors are to be used as current sources in a current-steering DAC it
is important that the transistors are well matched to obtain the correct bit weights
(within tolerable margins). Hence, it is desirable to manufacture transistors with
the same value of the parameters µ , C ox , V T , etc. However, when manufacturing
chips, variations in these parameters over the chip area occur, something we refer
to as component mismatch.
26
Modeling of static errors
I i = w i ⋅ I unit + ∆ i , (3.9)
where the first term, w i ⋅ I unit , is the desired current, and the second term, ∆ i , is
the deviation due to mismatch. This is equivalent to the model shown in Fig. 3.2,
where an error current source, ∆ i , is connected in parallel with the desired cur-
rent source, w i ⋅ I unit .
Ii wi Iunit Di
Ii = ∑
j∈J
I unit, j , (3.10)
i
i.e., the current source I i is a parallel connection of the unit current sources
I unit, j with indices j belonging to the set J i . We can rewrite (3.10) yielding
Ii = ∑
j∈J
I unit, j = w i ⋅ I unit + ∑ δ j .
j∈J
(3.11)
i i
∆i = ∑
j∈J
δj. (3.12)
i
If we consider two MOS transistors, M 1 and M 2 , on the same die, biased with
the same voltages and having (nominally) the same geometry, and compare the
corresponding drain currents I 1 and I 2 it has been shown that [15]
27
Modeling of Current-Steering DACs
A I2
Var(I 1 – I 2) = ------------- + S I2 ⋅ D 2 , (3.13)
W⋅L
where D is the distance between the centers of the transistors, and A I and S I are
process dependent constants. We see that for good matching properties (i.e.,
small variance) we should use transistors with large area ( W ⋅ L ) and keep them
as close as possible (small D ). Different yield estimation methods based on
(3.13) have been proposed in the literature [16, 17]. The use of a current-steering
DAC for efficient estimation of the parameters in (3.13) was presented in [18].
Another useful guideline for good matching properties, which is not captured in
(3.13), is to avoid covering the transistors with metal, since metal coverage
increases the variance [19].
Figure 3.3 Illustration of a parameter (e.g., oxide thickness) with circular distribution over
the wafer. Over a small area (e.g., a chip) the parameter value can be approxi-
mated with the first order Taylor expansion.
28
Modeling of static errors
where k x and k y are constants and x and y are the geometrical coordinates of the
current source with respect to a cartesian coordinate system. The matching error
δ j is given by
δ j( x , y ) = k x ⋅ x + k y ⋅ y . (3.15)
When implementing current-steering DACs, the unit current sources are often
placed in an array, as indicated in Fig. 3.4(a), where each square represents a unit
current source. A straightforward assignment of unit current sources to the differ-
ent current sources I i of a 6-bit binary weighted DAC is shown in Fig. 3.4(b),
where the upper half of the array is dedicated to the most significant bit (MSB),
the upper half of the remaining part of the array to MSB-1, etc., until the remain-
ing bits can be placed in the last row of the array.
y (b) y
5 5 5 5 5 5 5 5
5 5 5 5 5 5 5 5
5 5 5 5 5 5 5 5
x 5 5 5 5 5 5 5 5 x
4 4 4 4 4 4 4 4
4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3
2 2 2 2 1 1 0
(a) (b)
Figure 3.4 (a) Array of unit current sources and (b) unit current source assignment for a 6-
bit binary weighted DAC.
The way of assigning unit current sources indicated in Fig. 3.4(b) is known to
have poor properties in terms of suppressing graded matching errors. Different
layout techniques for suppressing the influence of graded matching errors have
been proposed in the literature [21, 22, 23]. However, we use this approach to
show how matching errors can affect the DAC performance. First of all, we gen-
eralize the concept illustrated in Fig. 3.4(b) for an N -bit binary weighted DAC
with the M least significant bits (LSBs) contained within the bottom row. For
comparison, the DAC in Fig. 3.4(b) has N = 6 and M = 3 . The array will then
have 2 M elements in the x -direction and 2 N – M elements in the y -direction.
Computing the values of the different ∆ i yields
29
Modeling of Current-Steering DACs
∆i = ∑
j∈J
δ j = ∑ kx ⋅ x j + ky ⋅ y j =
j∈J
i i
x y
= wi ⋅ k x ⋅
∑ -----j + k y ⋅ ∑ -----j = w i ⋅ ( k x ⋅ x i + k y ⋅ y i ) ,
w
j∈J i j ∈ J i
w
(3.16)
i i
where x j and y j are the coordinates of unit current source I unit, j , and x i and y i
are the average coordinates for the unit current sources assigned to bit i . We nor-
malize the coordinate axis such that the distance between adjacent unit current
sources are unity in both the x - and the y -direction, and we also normalize the
currents so that I unit = 1 . We have that
3
x i = 2 M – 1 + --- ⋅ 2 i for i = 0, …, M – 1 , (3.17)
2
x i = 0 for i = M , …, N – 1 , (3.18)
y i = – 2 N – M – 1 + 2 –1 for i = 0, …, M – 1 , (3.19)
and
3
y i = – 2 N – M – 1 + --- ⋅ 2 i – M for i = M , …, N – 1 . (3.20)
2
The parameter k x only affects the errors for the M LSBs. These errors are typi-
cally much less than the errors for the more significant bits (assuming that k x is
not very large compared with k y ) due to the factor w i in (3.16). With this as an
argument we set k x = 0 to simplify the calculations. This results in
∆ i = w i ⋅ k y ⋅ ( 2 –1 – 2 N – M – 1 ) =
= ( k y ⋅ ( 2 i – 1 – 2 i + N – M – 1 ) ) for i = 0, …, M – 1 (3.21)
and
∆ i = w i ⋅ k y ⋅ --- ⋅ 2 i – M – 2 N – M – 1 =
3
2
= k y ⋅ --- ⋅ 2 2 ⋅ i – M – 2 i + N – M – 1 for i = M , …, N – 1 .
3
(3.22)
2
30
Modeling of static errors
N–1 N–1
e(n) = ∑ bi(n) ⋅ I i – ∑ bi(n) ⋅ wi ⋅ I unit , (3.23)
i=0 i=0
i.e., the actual output signal minus the wanted output signal. Clearly,
N–1 N–1
e(n) = ∑ bi(n) ⋅ ∆i = ∑ ei(n) , (3.24)
i=0 i=0
where e i(n) = b i(n) ⋅ ∆ i . The power of this error signal, P e , can be expressed as
N–1
P e = Var(e(n)) = ∑ Var(ei(n)) + i∑≠ j Cov(ei(n), e j(n)) . (3.25)
i=0
If we assume that the probability that b i(n) = 1 and b i(n) = 0 are equal, i.e.,
1
P(b i(n) = 1) = P(b i(n) = 0) = --- (3.26)
2
we obtain
1 1 2
Var(b i(n)) = E(b i2(n)) – E 2(b i(n)) = --- – --- = --- ,
1
(3.27)
2 2 4
which yields
∆ i2
Var(e i(n)) = Var(∆ i ⋅ b i(n)) = ∆ i2 ⋅ Var(b i(n)) = ------ . (3.28)
4
The covariance terms in (3.25) are signal dependent, since
and can therefore not be expressed explicitly for an arbitrary signal. We shall in
the following assume that the covariance sum in (3.25) is zero (or small), which
is true if, e.g., the bits are mutually uncorrelated. With this assumption, P e
becomes
N–1
Pe = ∑ Var(ei(n)) . (3.30)
i=0
31
Modeling of Current-Steering DACs
4 –( M + 2 ) ⋅ T ⋅ k y2
P e = ------------------------------------- , (3.31)
105
where T is a number calculated according to
T = 70 ⋅ 2 M + N + 20 ⋅ 2 3M + N + 2 3 + 4N – 35 ⋅ 4 M – 28 ⋅ 4 2M – 35 ⋅ 4 N .
(3.32)
If we assume that the 2 3 + 4N = 8 ⋅ 2 4N term is dominating (which is true for rea-
sonable values of N and M , e.g., N = 14 and M = 6 ), we can approximate P e
with
8 ⋅ 2 4N 2 4N
P e ≈ --------------------------------
- ⋅ k = - ⋅ k y2 .
---------------------- (3.33)
16 ⋅ 105 ⋅ 2 2M y
210 ⋅ 2 2M
a2 2 2N – 2 2 2N
P s = ----- = --------------- = -------- (3.34)
2 2 8
and the power of the quantization noise is [7, 9]
1
P Q = ------ . (3.35)
12
Assuming that the quantization noise is uncorrelated with e(n) yields the follow-
ing expression for the SNDR
2( N + M )
Ps 105 ⋅ 2
2
-.
SNDR = ------------------- = --------------------------------------------------- (3.36)
Pe + PQ 4 ⋅ 2 4N ⋅ k y + 70 ⋅ 2 2M
To check the validity of (3.37) and (3.38) we have performed some simulations in
Matlab. A 14-bit DAC architecture has been used in the simulation and M has
been chosen to 7. A full-scale sinusoidal input was used, and matching errors
32
Modeling of static errors
were added to the bit weights according to (3.21) and (3.22). Simulated signal
spectra for k y = 10 –4 and k y = 10 –3 are plotted in Fig. 3.16(a) and (b), respec-
tively. In Fig. 3.16(c) we show simulated and calculated (using (3.36)) values of
SNDR vs. k y . The corresponding plots for SFDR, with the calculated values
given by (3.1), are shown in Fig. 3.5(d). In this example we conclude that simu-
lated and calculated values agree well, e.g., the simulated and calculated SNDR
has a maximum deviation of less than 2.5 dB even if rough approximations are
made. In [20] there was an extended analysis taking into account the correlation
between the input bits b N – 1 and b N – 2 , reducing the maximum deviation
between the curves to approximately 0.1 dB.
−5 −4
Output signal spectrum, k =10 Output signal spectrum, k =10
y y
0 0
−20 −20
−40 −40
PSD [dB]
PSD [dB]
−60 −60
−80 −80
−100 −100
−120 −120
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Normalized frequency Normalized frequency
(a) (b)
SNDR vs. ky SFDR vs. ky
90 120
simulated simulated
calculated calculated
110
80
100
70 90
60 80
SNDR [dB]
SFDR [dB]
70
50
60
40 50
40
30
30
20
20
10 −7 −6 −5 −4 −3 −2
10 −7 −6 −5 −4 −3 −2
10 10 10 10 10 10 10 10 10 10 10 10
ky ky
(c) (d)
Figure 3.5 Simulated signal spectra of 14-bit DAC with M = 7 for (a) k y = 10 –4 and (b)
k y = 10 –3 . The plots in (c) and (d) show simulated (dashed) and calculated
(solid) values of SNDR and SFDR vs. k y .
33
Modeling of Current-Steering DACs
∆ j ∈ N (0, σ) , (3.40)
where the different ∆ j are mutually uncorrelated. This way of modeling match-
ing errors will be used in Chapter 4 to analyze some methods for error compensa-
tion.
VDD
IunitX Runit/X
Vout(X)
RL
Figure 3.6 Equivalent model of current-steering DAC with current sources having finite
output resistance.
34
Modeling of dynamic errors
V DD
V out(X ) = I unit + ----------
X
- ⋅ ----------------------------- ⋅ R L , (3.41)
R unit RL
1 + ----------- ⋅ X
R unit
35
Modeling of Current-Steering DACs
Vbias
Iout Iout
(a) (b)
I
Rout Cout
cascode transistor
Vcasc Vout
Iout
Iout
(c) (d)
Figure 3.7 (a) Ideal current source, (b) single transistor PMOS implementation of a current
source, (c) single cascode PMOS implementation of a current source, and (d)
linearized model of the nonideal current source.
36
Modeling of dynamic errors
10
10
8
10
7
10
6
10
2 3 4 5 6 7 8
10 10 10 10 10 10 10
Frequency [Hz]
Figure 3.8 Output impedance vs. frequency for (a) a single transistor current source, and
(b) a single cascode current source.
switch-on resistance, as shown in Fig. 3.9(b), assuming that the switch transistor
operates in the linear region. This is a simplification, because the voltage over the
switch is varying with the output voltage of the DAC, and for some configura-
tions the switch transistor may even enter the saturated region. Thus, the linear-
ized model of the switch is a coarse approximation. However, it is good enough
to illustrate the effect of finite output impedance in the current sources. There is
also parasitic capacitance in the switch, but it is lumped into the parasitics of the
current source and output wire.
Iin
Iin
Rswitch
Q+ Q–
Iout+ Iout–
Iout+ Iout–
(a) (b)
Figure 3.9 (a) Example of a differential PMOS switch and (b) linearized model of the dif-
ferential switch.
37
Modeling of Current-Steering DACs
The output wires of the DAC should have zero impedance to reduce the internal
voltage drop. In reality, the wires contain resistive as well as capacitive and
inductive parts [29]. For high accuracy, a transmission line model or an RC-lad-
der network should be used as a model for the wire. However, we want to achieve
short simulation times, so a small number of circuit nodes is desired. Therefore,
we trade accuracy for a lower degree of complexity in the model. As an approxi-
mation we use a resistor and capacitor in parallel, i.e., a simple RC-ladder, as
shown in Fig. 3.10. The impedance includes the internal wire impedance as well
as the off-chip load.
( V DD – V out(t) ) ∂
I out(t) = I + ------------------------------------- – C out ⋅ V out(t) =
R out ∂t
V out(t) ∂
- – C out ⋅ V out(t)
= Ĩ – --------------- (3.42)
R out ∂t
where
V DD
Ĩ = I + ---------- . (3.43)
R out
38
Modeling of dynamic errors
VDD
I0 R0 C0 Ij Rj Cj IN1 CN1
RN1
Figure 3.11 Circuit schematic of the model of the nonideal differential current-steering
DAC.
Hence, the linearized model of a unit current source is equivalent to the circuit in
Fig. 3.12, where the current I unit is replaced with the slightly larger current Ĩ unit
and the output impedance connected to ground instead of V DD . This is the model
we will use from now on to eliminate terms including V DD from the calculations.
Ĩ unit
output impedance
Assume that we during an update period have a value of the digital input, X , such
that K of the current sources are connected to one of the analog outputs. For sim-
ple notation we let these current sources, I j , have indices j = 1, 2, …, K . The
circuit schematic for this specific configuration shown in Fig. 3.13.
The following relations between voltages and currents hold for this circuit
39
Modeling of Current-Steering DACs
I1 IK
V1 VK
Rs1 R1 C1 RsK RK CK
Iout1 IoutK
R C
Figure 3.13 Circuit schematic of one output during an update period for a certain value of
the input X .
V j(t) ∂
I outj(t) = I j – ----------- – C j ⋅ V i(t) for j = 1, 2, …, K , (3.44)
Rj ∂t
V j(t) – V (t)
- for j = 1, 2, …, K ,
I outj(t) = -------------------------- (3.45)
R sj
V (t )
I (t) = --------- , (3.46)
R
and
K
∂
I (t) = ∑ I outj(t) – C ⋅ ∂ t V (t) . (3.47)
j=1
∂
V i(t) = ------ ⋅ I j – V j(t) ⋅ ----- + ------- + V (t) ⋅ ------- .
1 1 1 1
(3.48)
∂t Cj R j R sj R sj
40
Modeling of dynamic errors
We let T denote the update period of the DAC. We are interested in finding the
values of the different nodal voltages at the end of the update period. For this pur-
pose we define the voltage vector
For simplicity we let the update period start at time t = 0 . We further define the
state vector
∂ ∂
V (t) = X(t) = A ⋅ X(t) + B ⋅ u(t) , (3.52)
∂t ∂t
where
-----
1
- + -------- ------
1 1
0 … 0
1 1
-------- ------
R 1 R s1 C 1 R s1 C 1
0 -----
1
- + -------- ------ …
1 1
0
1 1
-------- ------
R 2 R s2 C 2 R s2 C 2
A = … … … … …
(3.53)
and
41
Modeling of Current-Steering DACs
I1 I2 IK T
B = A⋅ V (0 –) + ------, ------, …, -------, 0 . (3.54)
C1 C2 CK
u(t) is a unit step used to apply the initial conditions given in B at time t = 0 . A
simple manipulation of (3.51) yields
Simulations
In this section we present simulation results to illustrate the effect of the nonideal
components introduced in the model. We consider a 14-bit binary weighted
architecture. The parameters used in the simulations are given in Table 3.1.
Fig. 3.14(a) and (b) show simulated PSD plots for single-ended and differential
outputs, respectively. The input is a full-scale, single-tone signal with signal fre-
quency f sig = 1.104 MHz and update frequency f u = 10 MHz .
Fig. 3.15 show the simulated PSD plots when the signal frequency is reduced to
f = 110.4 kHz , i.e., one tenth of the signal frequency used in Fig. 3.14. Com-
paring the two figures we find that the harmonic distortion is largely reduced
42
Modeling of dynamic errors
Parameter Value
Unit current 1.22 µA
Output resistance, unit current source 1 GΩ
Output capacitance, unit current source 5 fF
100 Ω , j = 0, …, 7
Switch resistance
100 ⋅ 2 7 – j Ω , j = 8, …, 13
Load resistance 70 Ω
Load capacitance 200 pF
Table 3.1 Model parameter values used in the simulations.
PSD for single−ended output signal PSD for differential output signal
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 3.14 Simulated (a) single-ended and (b) differential outputs for a full-scale single-
tone input with 10 MHz update frequency and 1.104 MHz signal frequency.
when the signal frequency is lowered. For the differential case the SFDR (deter-
mined by the third harmonic) is increased from 70 to 87 dB. We can also see that
the even order harmonics that appear in the single-ended spectra (the (b) figures)
are efficiently cancelled in the differential case.
In Fig. 3.16 we show how SFDR varies with signal frequency for a half-scale sin-
gle-tone signal with update frequency f s = 10 MHz . This behavior is typical
for current-steering DACs [10, 32, 33], compare, e.g., with the measurement
results presented in Chapter 5.
43
Modeling of Current-Steering DACs
PSD for single−ended output signal PSD for differential output signal
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 3.15 Simulated (a) single-ended and (b) differential outputs for a full-scale single-
tone input with 10 MHz update frequency and 110.4 kHz signal frequency.
90
70
85
SFDR [dB]
SFDR [dB]
65 80
75
60
70
55
5 6 5 6
10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)
Figure 3.16 Simulated SFDR as a function of signal frequency for (a) single-ended output
and (b) differential output. The update frequency is 10 MHz, and the signal is a
full-scale single-tone signal.
44
Modeling of dynamic errors
IX
V(t)
R CX
This is a single pole system with the output voltage, V (t) , given by
t – t0
– ------------
V (t) = V (t 0) + ( I X ⋅ R – V (t 0) ) ⋅ 1 – e RC X , (3.56)
where I X is the output current for the input code X , nominally I unit ⋅ X , and C X
is the signal dependent parasitic load. In discrete time we are interested in the
value of the output at multiples of the update period T , i.e.,
T
– -----------
Y (n) = V (nT ) = Y (n – 1) + ( I X ⋅ R – Y (n – 1) ) 1 – e RC X , (3.57)
where t 0 has been chosen as ( n – 1 )T . From (3.57) it can be seen that Y (n) is
depending on Y (n – 1) , which in turn is dependent on Y (n – 2) , etc. Thus, the
nonlinearity is dynamic and has infinite memory. We can however define a static
parameter describing the error, namely the relative step error
Y (n) – Y (n – 1)
e rel(X ) = 1 – ------------------------------------ (3.58)
Y X – Y (n – 1)
where Y X is the settled output value for input code X , Y X = I X ⋅ R for the sim-
ple model in Fig. 3.17. Combining (3.57) and (3.58) yields
45
Modeling of Current-Steering DACs
T
– -----------
RC X
e rel(X ) = e (3.59)
for the simple model in Fig. 3.17. This expression can be interpreted as that the
DAC has a signal dependent settling time constant τ X = RC X .
Model structure
Based on the simple model in the previous section, we propose a low-complexity
model to be used for estimation of the output in a current-steering DAC with non-
zero output capacitance in the current sources, as discussed in Sec. 3.4.1. We start
by normalizing the output such that
YX = X , (3.60)
i.e., we let the static gain of the DAC be unity. We rewrite (3.57)-(3.59) yielding
z1
Figure 3.18 Proposed model.
Y (n) – Y (n – 1)
e rel(X (n)) = 1 – ------------------------------------- (3.62)
Y X (n) – Y (n – 1)
46
Modeling of dynamic errors
−3 erel vs X −3 erel vs X
x 10 x 10
10 10
5 5
Relative error
Relative error
0 0
−5 −5
0 2000 4000 6000 8000 10000 12000 14000 16000 0 2000 4000 6000 8000 10000 12000 14000 16000
Input code Input code
(a) (b)
Figure 3.19 The relative step error, e rel , plotted as a function of the input code X (a) with-
out and (b) with the estimated values using the low-complexity model.
A block diagram of a circuit that calculates (3.63) is shown in Fig. 3.20. Again
we have set the static gain of the DAC to unity, replacing Y X with X in (3.63).
The lookup table may be addressed by only a few MSBs of X (n) and X (n – 1) to
keep the size of the lookup table small, as discussed above.
47
Modeling of Current-Steering DACs
X(n)
z1 lookup erel(n) Y(n)
table
z1
Figure 3.20 Block diagram for the modified approach.
48
Combined models
Single−ended PSD for low−complexity model Differential PSD for low−complexity model
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 3.21 (a) Single-ended and (b) differential outputs for a full-scale single-tone input
with 10 MHz update frequency and 110.4 kHz signal frequency simulated with
the low-complexity model.
85
85
SFDR [dB]
SFDR [dB]
80
80
75
75
70 70
5 6 5 6
10 10 10 10
Frequency [Hz] Frequency [Hz]
(a) (b)
Figure 3.22 Simulated SFDR as a function of signal frequency for differential output signals
with full-scale amplitude. The plot in (a) shows the results for the low-complex-
ity model, whereas (b) shows the results for the state-space model for compari-
son.
49
Modeling of Current-Steering DACs
50
4 Compensation and
Correction of Errors
It can be hard to meet a DAC design specification using a straightforward imple-
mentation. In this chapter we present some ideas on how to compensate for the
errors, in order to improve the linearity of the circuits. So called predistortion is a
family of techniques widely used in telecommunication circuits, mostly RF
power amplifiers, to enhance the linearity and, thereby, improve the quality of the
communication link. The idea is to use models of the nonlinearity and modify the
input signal in order to get the desired output signal. Other techniques, e.g.,
dynamic element matching (DEM), rely on randomization in order to change the
spectral properties of the nonlinearities from spurious tones to (white or shaped)
noise.
The presentation of the different techniques given in this chapter has been
divided into two main groups, compensation and correction of static errors,
which is presented in Sec. 4.1, and compensation and correction of dynamic
errors, which is presented in Sec. 4.2.
51
Compensation and Correction of Errors
Basic idea
The idea behind the proposed calibration method is to remove the DNL error in
the transition from X = [ 0, 1, 1, …, 1 ] to X = [ 1, 0, 0, …, 0 ] . This error is
often the dominating one in binary coded DACs due to the large amount of
switches turning on and off. To achieve this, the MSB current, I N – 1 , is substi-
tuted with a current
N–2
Ĩ N – 1 = I unit + ∑ Ij, (4.1)
j=0
with the intention of using the currents from the less significant bits and one addi-
tional unit current source as reference for the calibration. The DNL and INL for a
12-bit binary coded DAC with random matching errors having a standard devia-
tion of 1% are shown in Fig. 4.1 and Fig. 4.2, respectively. Values without cali-
bration are given in the (a) figures, whereas values with calibration are given in
the (b) figures. In Fig. 4.2(a) we can observe a problem with the binary weighted
architecture in that its transfer characteristics is not always monotone. In this
case,
N–2
IN – 1 < ∑ Ij. (4.2)
j=0
52
Compensation and correction of static errors
0.5 0.5
0 0
−0.5 −0.5
−1 −1
DNL
DNL
−1.5 −1.5
−2 −2
−2.5 −2.5
−3 −3
−3.5 −3.5
−4 −4
0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
Input code Input code
(a) (b)
Figure 4.1 DNL for binary weighted DAC (a) without and (b) with calibration of the MSB.
2 2
1 1
INL
INL
0 0
−1 −1
−2 −2
−3 −3
0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
Input code Input code
(a) (b)
Figure 4.2 INL for binary weighted DAC (a) without and (b) with calibration of the MSB.
0 0
−20 −20
−40 −40
PSD [dB]
PSD [dB]
−60 −60
−80 −80
−100 −100
−120 −120
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Normalized frequency Normalized frequency
(a) (b)
Figure 4.3 PSD plot for full-scale single-tone signals (a) without and (b) with calibration of
the MSBs.
53
Compensation and Correction of Errors
Proposed implementation
A proposed circuit for the calibration technique is shown in Fig. 4.4. The calibra-
tion phase is shown in Fig. 4.4(a). A current mirror is utilized to construct the dif-
ference ∆I N – 1 = I N – 1 – Ĩ N – 1 , and a current memory is set to hold the current
∆I N – 1 . During the operation phase, shown in Fig. 4.4(b), the current memory is
utilized to subtract the current ∆I N – 1 from I N – 1 , yielding the total output cur-
rent Ĩ N – 1 .
DIN1
current
current memory mirror
(a)
IN1 IN2 I0
DIN1
(b)
Figure 4.4 Proposed circuit solution for the calibration technique during (a) the calibration
phase and (b) the operation phase.
54
Compensation and correction of static errors
W out
I out, j = ----------- ⋅ I in (4.3)
W in
provided that the output transistors are operating in the saturated region and that
the transistors are perfectly matched. The transistors have the same length. W out
denotes the width of the output transistors and W in is the width of the input tran-
sistor. In the following we discuss how current mirrors can be utilized for biasing
current sources and how the matching properties are affected.
In Fig. 4.6(a) we show an example where the current-source array is made up of
the output transistors of a single multiple-output current mirror. Graded parame-
ter variations, discussed in Sec. 3.3.1, cause the output currents from the different
current sources to vary over the array. The matching of output currents can be
improved by partitioning the array of current sources into smaller parts, letting
each such part be an individual multiple-output current mirror, as indicated in
Fig. 4.6(b). This is because the bias voltage ( V b in Fig. 4.5) is set individually for
55
Compensation and Correction of Errors
Vb
each current mirror. If the input transistors are located close to the output transis-
tors, the bias voltage is set to give the proper output current for the values of tran-
sistor parameters in the proximity of the input transistors, reducing the current
variations over the array due to graded parameter mismatch, compared with the
approach in Fig. 4.6(a). The biasing strategy has no effect on the influence of ran-
dom matching errors, so this problem remains.
56
Compensation and correction of static errors
Ibias Ibias,1
Ibias,2
Ibias,K
(a) (b)
Figure 4.6 Current-source arrays with (a) global and (b) distributed bias.
suitable in audio applications where the signal bandwidths are low (compared
with, e.g., DSL applications) and a large amount of oversampling can be
afforded.
Generalized DEM
A general DEM DAC is shown in Fig. 4.7. The input, X 1(n) , is converted into a
thermometer coded signal
such that
57
Compensation and Correction of Errors
and for each j there is exactly one i such that (4.6) holds. Which t j(n) is mapped
on what u j(n) is determined for each update instant n by the control signal r(n) .
Each u i(n) is an input to a one bit DAC with nominal gain w and an associated
matching error ∆ i . The outputs from all one bit DACs are summed, yielding the
total output
K K K
Y (n) = ∑ ( w + ∆i ) ⋅ ui(n) = w ⋅ ∑ ui(n) + ∑ ∆i ⋅ ui(n) =
i=1 i=1 i=1
K
= w ⋅ X 1(n) + ∑ ∆i ⋅ ui(n) , (4.7)
i=1
where the first term in is the desired output, and the second term is the error due
to mismatch. Without scrambling, the error becomes
K
∑ ∆i ⋅ t i(n) (4.8)
i=1
and harmonic distortion occurs because the resulting error is periodic. This is not
true in DEM DACs, provided that the control signal r(n) is properly chosen. The
scrambling ensures that different errors occurs for the same input code at differ-
ent update instants, a property which is used to transform harmonic distortion
into noise. The overall SNDR is improved if oversampling is used, because the
out-of-band noise can be filtered out.
A straightforward implementation of the digital circuits required in DEM may
result in high hardware complexity. There is a large amount of research on DEM
circuit techniques with reduced hardware complexity presented in the literature
[41, 42]. Some of these techniques are overviewed in the following sections. For
comparisons of different networks used for DEM realizations, see, e.g., [43, 44].
58
Compensation and correction of static errors
binary-to-thermometer encoder
scrambler
X1(n)
1-bit combined
X2(n)
X3(n)
r(n)
DACs output
the input, whereas the value of the MSB of the input is assigned to all individual
bits of out 2 . When r i = 0 the MSB of the input is instead mapped on the bits of
out 1 , and the M – 1 LSBs of the input are mapped on out 2 . The number of out-
put bits for each switch is linearly decreasing with decreasing layer index, and
the tree is terminated with the switching layer having 1-bit outputs. These 1-bit
signals constitute a scrambled, thermometer coded word that is used to control 1-
bit DACs, whose outputs are summed yielding the total output for the DEM
DAC.
An additional bit with the same weight as the LSB, and fixed value ‘0’, is added
to the binary weighted input to give the LSBs the same total weight as the MSB.
This is required because there is an even number of output bits from the switch-
ing tree, whereas a thermometer coded word has an odd number of bits. This
results in that at least one of the output bits from the tree will have the value ‘0’.
59
Compensation and Correction of Errors
layer 1
2
switch
layer N1 r1(n)
N
switch
to 1-bit DACs
layer N
N N+1 rN1(n)
X(n) switch
rN(n) N
0 switch
rN1(n)
2
switch
r1(n)
1
u i(n) = ---- ⋅ X 1(n) + z i(n) , (4.9)
K
where K is the number of 1-bit signals and z i(n) is a signal with very little power
within the signal band. The total output is
K K
∑ ∑ ---K- ⋅ ( w + ∆i ) ⋅ zi(n) =
1 1
Y (n) = ---- ⋅ ( w + ∆ i ) ⋅ X 1(n) +
K
i=1 i=1
K K
= X 1(n) ⋅ w ⋅ 1 +
∑
∆ i +
∑
( w + ∆i ) ⋅ zi . (4.10)
i=1 i=1
60
Compensation and correction of dynamic errors
The first term in (4.10) is the desired output (with a slightly different gain com-
pared with (4.7)). The second term is the error, e(n) , due to mismatch. Moreover
K K
X 1(n) = ∑ ui(n) = X 1(n) + ∑ zi(n) , (4.11)
i=1 i=1
yielding
K
∑ zi(n) = 0 . (4.12)
i=1
Since every z i has very little power within the signal band, so does e(n) . The
construction of the different u i(n) is similar to delta-sigma modulation (which is
outlined in Sec. 4.2.2) and is nontrivial. Several circuit solutions have been pro-
posed (see, e.g., [24, 45, 47, 48]), and there are also solutions where the different
u i(n) are multi-bit words.
61
Compensation and Correction of Errors
A+
X1(n)
DAC1
A
X(n) r(n)
Vdump
A+
X2(n) I2 I1
DAC2
A RL RL
where X (n) is the input to the whole system and r(n) is a control signal. The
input, X 2(n) , to DAC2 is similarly given by
62
Compensation and correction of dynamic errors
where X max is the maximum value of X for the given code, i.e., with all individ-
ual bits b i = 1 . If the input code is binary weighted, then X max – X (n) corre-
sponds to X (n) with all individual bits inverted, represented by the inverter
symbol in Fig. 4.9. The control signal must not cause any of the signals X 1 and
X 2 to overflow, since this causes distortion. Thus, the DAC inputs need to fulfill
0 ≤ X i ≤ X max , (4.16)
yielding
With ideal DACs (DAC1 and DAC2) we obtain the following expressions for the
output currents
N–1
I1 = Iu ∑ wi ⋅ bi(n) + I u ⋅ r(n) and
i=0
N–1
I2 = Iu ∑ wi ⋅ bi(n) + I u ⋅ r(n) . (4.18)
i=0
Hence, ideally we have the same differential output current as for a single ideal
differential current-steering DAC, because the control signals cancel. Of course,
in reality the control signal will affect the two outputs differently due to any non-
linearity in the transfer functions of the DACs, and the cancellation will not be
perfect.
We also consider the common-mode signal
N–1
I1 + I2
I cm = ---------------
2
= Iu ∑ wi + I u ⋅ r(n) . (4.20)
i=0
Hence, the control signal is added to the common-mode signal. One factor that
limits the possibly useful control signals is, therefore, the common-mode rejec-
tion of the following circuitry. Another factor is, as mentioned earlier, how the
control signal is distorted between the two DAC outputs.
63
Compensation and Correction of Errors
a ⋅ ψ(n) , (4.21)
where a is a scaling constant and ψ(n) is white noise with a rectangular distribu-
tion
DC level minimization
We propose a second method based on the proposed redundant architecture in
Fig. 4.9 that aims at reducing the parasitic load associated with an output termi-
nal. Looking at the circuit models presented in Sec. 3.4 we see that each current
64
Compensation and correction of dynamic errors
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [MHz] Frequency [MHz]
(a) (b)
Figure 4.10 PSD plots for simulated outputs from redundant DAC with matching errors (a)
without dither signal and (b) with dither signal having amplitude a = 2 13 .
98
96
94
92
SFDR [dB]
90
88
86
84
82
80
4 6 8 10 12 14
log2(a)
65
Compensation and Correction of Errors
Here we present some simulation results using the model presented in Sec. 3.4.1.
The simulated DAC is 14-bit binary-weighted, and the model parameters are the
same as in Table 3.1 on page 43. The current sources are assumed to be perfectly
matched, but matching errors can also be added in the model, as discussed in
Sec. 3.5. PSD plots of the simulated DAC outputs using half-scale sinusoidal
inputs are shown in Fig. 4.12. The PSD for a single differential DAC is shown in
Fig. 4.12(a). PSDs for the proposed redundant architecture, with the dump termi-
nal connected to a DC voltage source of 0.5 V, are shown in Fig. 4.12(b) and (c).
The PSD for r = 0 is shown in Fig. 4.12(b) and the PSD for r = – 2 12 (yielding
the lowest possible DC level without overflow) is shown in Fig. 4.12(c). A Com-
parison of Fig. 4.12(a) and (b) indicates that switching a current source between
an output node and a silent node has favorable effects on the distortion, compared
with switching between two output nodes, since the distortion is 5 dB lower in
Fig. 4.12(b). This also agrees with some single-ended measurements performed.
The spurious tones are reduced even more, resulting in an improvement in the
SFDR by approximately 7 dB when the DC level is reduced, as is the case in
Fig. 4.12(c). Further measurements have to be performed before any conclusions
can be made regarding the application to real DAC circuits.
and
Boundary condition (4.23) ensures that overflow is avoided, and ∆r max is chosen
such that the common mode variations can be sufficiently rejected in the follow-
ing circuitry. The previous approach with a constant r corresponds to
∆r max = 0 . The boundary conditions should hold for all n , so we need to keep
track of some future samples of X , otherwise we risk choosing r(n) too small to
fulfil both (4.23) and (4.24) for some future sample. However, since we do not
need to keep track of all future samples, which is the case when using a constant
r , this approach is more suitable for implementation.
66
Compensation and correction of dynamic errors
−20
−40
PSD [dB/Hz]
−60
−80
−100
−120
0 1 2 3 4 5
Frequency [Hz] 6
x 10
(a)
PSD for output signal, r=0 PSD for output signal, r=−4096
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(b) (c)
Figure 4.12 PSD plots of simulated DAC outputs from (a) a conventional DAC, and the pro-
posed redundant DAC with (b) r = 0 and (c) r = – 2 12 . The input is a half-
scale sinusoid with signal frequency 2.208 MHz and update frequency 10 MHz.
In Fig. 4.13(a) and (b) we show simulated 3-tone PSD plots for a single DAC and
for the proposed redundant DAC with r(n) chosen as above, respectively. The
simulation model is a 14-bit binary weighted version of the one presented in
Sec. 3.4.1. Parameter values are given in Table 3.1 on page 43, and ∆r max = 20 .
The peak-to-peak value for this particular input is approximately 10600. The
largest distortion peak is reduced from –72 to –82 dB.
67
Compensation and Correction of Errors
PSD for simulated output signal, r=0 PSD for simulated output signal, ∆r =10
max
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 4.13 PSD for 3-tone signal (a) without and (b) with common-mode level reduction.
ever, an ideal quantizer is a poor model of a real data converter. Reference levels
will not be evenly spaced due to, e.g., component mismatch. Other dynamic and
static errors also occur, as was discussed in Chapter 3.
In this section we propose a linearization technique based on delta-sigma modu-
lation, primarily intended for DACs, which also spectrally shapes nonlinearities
arising from other sources than the quantization process. The application, how-
ever, is not limited to DACs, but can be applied to any kind of oversampled sys-
tem. The technique presented here was first proposed in [56].
(a) (b)
Figure 4.14 (a) General delta-sigma modulator and (b) delta-sigma modulator with quanti-
zer modeled as an added error signal.
68
Compensation and correction of dynamic errors
X(n) Y(n)
Q
e(n)
H(z)
Figure 4.15 Error feedback modulator.
STF(z) = 1 (4.26)
and
69
Compensation and Correction of Errors
−20
−40
PSD [dB]
−60
−80
−100
−120
0 0.1 0.2 0.3 0.4 0.5
Normalized frequency
(a)
6−bit output signal from modulator 6−bit single−tone signal
0 0
−20 −20
−40 −40
PSD [dB]
PSD [dB]
−60 −60
−80 −80
−100 −100
−120 −120
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Normalized frequency Normalized frequency
(b) (c)
Figure 4.16 Spectra for (a) 14-bit single-tone signal, (b) delta-sigma modulated signal using
5th order modulator with a 6-bit quantizer, and (c) 6-bit single-tone signal.
70
Compensation and correction of dynamic errors
X(n) DAC
Q model
e(n)
H(z)
Figure 4.17 Block diagram of the proposed distortion shaping device.
Simulations
A suitable candidate for the model in the feedback loop to compensate for errors
due to finite output impedance is the low-complexity model presented in
Sec. 3.4.2. Here some simulation results for linearization of the DAC model in
Sec. 3.4.1 are presented. The simulation parameters are the same as in Table 3.1
on page 43, except that the output resistance in the unit current sources has been
increased a factor 10 to reduce the static nonlinearity, which is not included in the
low-complexity model. The parameters for the low-complexity model are the
same as those used in the simulations in Sec. 3.4.2. The filter H (z) is the same as
in the delta-sigma simulation in Fig. 4.16. The method is illustrated in by the
simulation plot in Fig. 4.18 where the dominating second harmonic is suppressed
with 20 dB, to the cost of increased high frequency distortion. There is one spuri-
ous tone within the signal band that increases, so the SFDR improvement is only
11 dB.
71
Compensation and Correction of Errors
Output signal wihtout distortion shaped Output signal with distortion shaped
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 4.18 PSD plots for simulated outputs (a) without and (b) with distortion shaping.
Measurements
Measured PSD plots for single-ended outputs from the DAC presented in
Sec. 5.5 are shown in Fig. 4.19. Fig. 4.19(a) shows the spectrum using an uncom-
pensated input, whereas the input used in Fig. 4.19(b) has been processed with
the distortion shaping technique presented in this section. The model used for
estimation of the error is a simple version of the low-complexity model presented
in Sec. 3.4.2, namely
Referring to the block diagram of the model shown in Fig. 3.18, the lookup table
only contains two values, a and 0, and is addressed by the MSB of the input to
the DAC. The resulting SFDR improvement is approximately 12 dB, which gives
a strong indication that the model structure presented in Sec. 3.4.2 is suitable for
its purpose, i.e., to estimate dynamic errors in order to compensate for them.
72
Compensation and correction of dynamic errors
(a) (b)
Figure 4.19 PSD plots for measured outputs from a DAC (a) with and (b) without the distor-
tion shaped.
Predistortion block
The purpose of the predistortion block is to, for a given input X (n) , find a modi-
fied input, X p(n) , such that Y (n) ≈ X (n) . To accomplish this we use the low-com-
plexity DAC model presented in Sec. 3.4.2. In accordance with Sec. 3.4.2 we use
as an estimate of the output from the current-steering DAC and normalize the
output such that the static gain of the DAC is unity. To find the appropriate pre-
distorted signal we replace X with X p in (4.29), yielding
73
Compensation and Correction of Errors
X (n) – X (n – 1)
X p(n) = X (n – 1) + ------------------------------------------------------------ . (4.32)
1 – e rel(X p(n), X p(n – 1))
In Sec. 3.4.2 only a few MSBs of X are used to estimate e rel . Thus, if we assume
that the correction introduced by the predistortion is small, i.e., X p(n) – X (n) is
small, it is reasonable to approximate e rel(X p(n), X p(n – 1)) with
e rel(X (n), X (n – 1)) . We can then make a nonrecursive definition of X p(n)
X (n) – X (n – 1)
X p(n) = X (n – 1) + ------------------------------------------------------- =
1 – e rel(X (n), X (n – 1))
= X (n – 1) + p rel(X (n), X (n – 1)) ⋅ ( X (n) – X (n – 1) ) , (4.33)
where
1
p rel(X (n), X (n – 1)) = ------------------------------------------------------- . (4.34)
1 – e rel(X (n), X (n – 1))
Similarly to the error estimation presented in Sec. 3.4.2, p rel can be stored in a
lookup table. Further, the lookup table may be addressed with only a few MSBs
of X (n) and X (n – 1) to keep the size of the lookup table small. We propose a
predistorter circuit according to Fig. 4.20.
X(n)
z1 lookup prel(n) Xp(n)
table
Figure 4.20 Proposed predistorter circuit.
Simulation results
In this section we present results from simulations where the predistortion tech-
nique is used to linearize the state-space model of the current-steering DAC pre-
sented in Sec. 3.4.1. The simulation parameters are given in Table 3.1 on page
43, and the parameters for the predistortion block are the same as in the simula-
tions in Sec. 3.4.2.
74
Compensation and correction of dynamic errors
The PSD plots for outputs without and with predistorted inputs are shown in
Fig. 4.21(a) and (b), respectively. The input is a half-scale sinusoid with signal
frequency 1.104 MHz and the update frequency is 10 MHz. Full-scale inputs
cannot be used since the predistorted inputs are likely to overflow in that case. In
Fig. 4.22 we have increased the signal frequency to 2.208 MHz. For both fre-
quencies it can be concluded that the predistortion reduces the second harmonic
to a large extent and suppresses the other harmonics below the noise floor. SFDR
is plotted as a function of signal frequency in Fig. 4.23. For this example the
SFDR at low frequencies is preserved over the whole frequency range when the
predistortion technique is used.
PSD for nonpredistorted input signal PSD for predistorted input signal
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 4.21 PSD plots for outputs (a) without and (b) with predistortion. The input is a half-
scale sinusoid with signal frequency 1.104 MHz, and the update frequency is
10 MHz.
PSD for nonpredistorted input signal PSD for predistorted input signal
0 0
−20 −20
−40 −40
PSD [dB/Hz]
PSD [dB/Hz]
−60 −60
−80 −80
−100 −100
−120 −120
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [Hz] 6 Frequency [Hz] 6
x 10 x 10
(a) (b)
Figure 4.22 PSD plots for outputs (a) without and (b) with predistortion. The input is a half-
scale sinusoid with signal frequency 2.208 MHz, and the update frequency is
10 MHz.
75
Compensation and Correction of Errors
80
75
SFDR [dB]
70
65
60
55
50 5 6
10 10
fsig [Hz]
Figure 4.23 Simulated SFDR as a function of signal frequency with and without predistor-
tion.
76
5 Current-Steering DAC
Implementations
The most common type of DAC for high-speed applications, e.g., broadband
communication, is the current-steering. The implementation of three different 14-
bit current-steering DACs in standard digital CMOS processes is discussed in this
chapter, together with the measurement results.
77
Current-Steering DAC Implementations
NMOS transistors
A commonly used symbol for an NMOS transistor is shown in Fig. 5.1(a). The
labels G, S, D, and B denote the gate, source, drain, and bulk terminal, respec-
tively. We have also indicated the gate-source voltage, V GS , the drain-source
voltage, V DS , and the bulk-source voltage, V BS . The drain current, I D is also
indicated in the figure. In most processes, the bulk terminals of all NMOS tran-
sistors are connected to ground, and the simplified symbol in Fig. 5.1(b), without
the bulk terminal explicitly shown, can be used instead.
ID
D
G B VDS
VBS
VGS S
(a) (b)
Figure 5.1 (a) Symbol for an NMOS transistor, and (b) simplified symbol for an NMOS
transistor with the bulk terminal connected to ground.
Equations (5.1), (5.2), and (5.3) model the relationships between the drain cur-
rent and the different voltages for a long channel NMOS transistor in the cut-off,
linear, and saturated region, respectively.
V GS < V T , n ⇒
ID = 0 , (5.1)
V GS ≥ V T , n , V DS < V GS – V T , n = V eff ⇒
V DS 2
= µ 0, n C ox ----- V eff V DS – ---------- ,
W
ID (5.2)
L 2
and
V GS ≥ V T , n , V DS ≥ V GS – V T , n = V eff ⇒
µ 0, n C ox W 2
I D = ------------------- ----- V eff ( 1 + λ ( V DS – V eff ) ) . (5.3)
2 L
78
CMOS processes
V T , n = V T , n, 0 + γ ( 2Φ F – V BS – 2Φ F ) , (5.4)
PMOS transistors
Symbols for PMOS transistors are shown in Fig. 5.2. In Fig. 5.2(a) the four ter-
minal symbol is shown, and the corresponding simplified symbol where it is
implied that the bulk terminal is connected to the power supply voltage ( V DD ) is
shown in Fig. 5.2(b).
VSG S
VSB
G VSD
B
ID D
(a) (b)
Figure 5.2 (a) PMOS transistor symbol and (b) simplified PMOS transistor symbol.
Equations (5.5), (5.6), and (5.7) model a long channel PMOS transistor in the
cut-off region, linear region, and saturated region, respectively
V SG < V T , p ⇒
ID = 0 , (5.5)
V SG ≥ V T , p , V SD < V SG – V T , p = V eff ⇒
V SD 2
I D = µ 0, p C ox ----- V eff V SD – ---------
- ,
W
(5.6)
L 2
and
79
Current-Steering DAC Implementations
V SG ≥ V T , p , V SD ≥ V SG – V T , p = V eff ⇒
µ 0, p C ox W 2
I D = ------------------- ----- V eff ( 1 + λ ( V SD – V eff ) ) . (5.7)
2 L
The threshold voltage for PMOS transistors is negative. Due to the body effect
we get
V T , p = V T , p, 0 – γ ( 2Φ F – V SB – 2Φ F ) . (5.8)
∂I D W
gm = ≈ 2µ 0, n C ox ----- I D , (5.9)
∂ V GS L
∂I D
g ds = ≈ λI D , (5.10)
∂ V DS
and
∂I D γ gm
g mbs = ≈ ------------------------------------ . (5.11)
∂ V BS 2 2Φ – V
F BS
The corresponding small signal equivalent model for PMOS transistors is shown
in Fig. 5.3(b). The parameters are given by
∂I D W
gm = ≈ 2µ 0, p C ox ----- I D , (5.12)
∂ V SG L
80
Analog DAC building blocks
G D S
id
vgs vds vsg vds
S G id D
gmvgs gmbsvbs gds gmvsg gmsbvsb gds
(a) (b)
Figure 5.3 Small signal equivalent model of (a) an NMOS transistor and (b) a PMOS tran-
sistor.
∂I D
g ds = ≈ λI D , (5.13)
∂ V SD
and
∂I D γ gm
g msb = ≈ ------------------------------------ . (5.14)
∂ V SB 2 2Φ – V
F SB
81
Current-Steering DAC Implementations
1
R out = ---------------------- < ∞ , (5.15)
g ds, source
1
Z out = ----------------------------------------- , (5.16)
g ds, source + sC out
where the output capacitance C out = C p, 1 and g ds, source is the g ds for the tran-
sistor.
Cp,1
Vbias source
transistor
Vbias Cp,1
Cp,2
Vcasc cascode
transistor
Iout
Iout
(a) (b)
Figure 5.4 PMOS current sources (a) without and (b) with cascode transistor.
g m, casc 1
R out ≈ ----------------- ⋅ ---------------------- , (5.17)
g ds, casc g ds, source
where g m, casc and g ds, casc are small signal parameters for the cascode transistor
and g ds, source is the g ds for the source transistor. The output resistance is
increased, since g m, casc » g ds, casc . The cascode transistor reduces the voltage
variations over the parasitic capacitance C p, 1 , compared with the single transis-
82
Analog DAC building blocks
tor case. Hence, the influence of C p, 1 on the bandwidth of the current source is
reduced. There are, however, also parasitic capacitance in the output node of the
cascode transistor. The output impedance is approximately
1
Z out ≈ ------------------------------------------------------------ , (5.18)
g ds, casc
g ds, source ----------------- + sC out
g m, casc
where in this case C out ≈ C p, 2 . For higher frequencies where the capacitive part
dominates, then Z out ≈ 1 ⁄ ( sC out ) for both type of current sources. C out is
roughly a parallel connection between the gate-drain and the drain-bulk capaci-
tance for the transistor connected to the output node. The value of the parasitic
capacitance is mainly determined by the width of the transistor. If the width is of
the same order in both the cascoded and the single transistor case, yielding
C p, 1 ≈ C p, 2 , we have not gained very much in the range of frequencies where
the capacitive part is dominating. A part of the output capacitance is also contrib-
uted to by the metal wires connected to the current sources. This part of the
capacitance is unaffected by the choice of current source. In the next section we
show that the cascoding effect can also be achieved with the switch. The conclu-
sions are that for higher frequencies we do not gain very much by introducing
cascode transistors but we loose in possible output swing since we need to keep
all transistors in saturation (the output resistance is severely deteriorated if the
transistors enter the linear region). Therefore, the cascode transistors have been
omitted in the later DAC implementations.
Iin Iin
Q+ Q Q+ Q
I+ I I+ I
(a) (b)
Figure 5.5 CMOS implementations of (a) a differential PMOS switch and (b) a differential
NMOS switch.
83
Current-Steering DAC Implementations
It is important that the signals controlling the switches are such that the two tran-
sistors in the switch are not simultaneously completely turned off, in which case
charge will be assembled at the output node of the current source and give rise to
large transients once one of the transistors is turned back on.
Assume that the PMOS type differential switch is used. In steady-state, when one
of the transistors is conducting and the other turned off, the conducting switch
transistor acts as a cascode transistor (compare with Fig. 5.4(b)). In order to have
short switching times and low channel charge injection from the switches it is
desired to use transistors with small area in the switches. A large g m ⁄ g ds is
required to get a good cascoding effect. Since g m ⁄ g ds is proportional to W ,
there is a trade-off between switch area and cascoding effect. Moreover, when the
output voltage exceeds the absolute value of the threshold voltage, the switch
transistor enters the linear region of operation and the g m ⁄ g ds ratio is deterio-
rated. Even if the cascoding effect of the switch is limited, it is better to use
PMOS switches than NMOS switches. Similarly, NMOS switches are preferred
when NMOS current sources are used.
84
Digital DAC building blocks
response of the DAC, as shown in Fig. 5.6. This can occur because in the transi-
tion there are individual bits making transitions from 0 to 1 , at the same time as
there are other bits making transitions from 1 to 0 .
15
14
13
12
11
10
output level
9
8
7
6
5
4
3
2
1
n
update instant
Figure 5.6 Large glitch due to poor internal timing in a 4-bit binary weighted DAC. The
desired transition is from 7 to 8.
Y (X 1) = ∑ wi = ∑ wi + ∑ w i = Y (X 2) + ∑ wi (5.19)
i=1 i=1 i = X2 + 1 i = X2 + 1
and all w i > 0 . As discussed in Sec. 4.1.1, a binary coded converter is not neces-
sarily monotone.
85
Current-Steering DAC Implementations
0 0 0 0 0 0 1 0 0 0 0 0 0 1 1
0 1 0 0 1 0 1 1 0 0 0 0 1 1 1
1 0 0 0 0 0 1 1 1 1
1 0 0 1 1
1 0 1 0 0 1 1 1 1 1
1 1 1 1 1
1 1 0 0 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
(a) (b)
b3 b2 b1 b0 t415 t414 t413 t412 t411 t410 t49 t48 t47 t46 t45 t44 t43 t42 t41
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1
0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1
0 1 0 1 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1
0 1 1 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1
0 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1
1 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1
1 0 0 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
1 0 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1
1 0 1 1 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1
1 1 0 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
(c)
Figure 5.7 Truth tables for (a) 2-3 encoder and (b) 3-7 encoder, and (c) 4-15 encoder.
Other types of codes that falls somewhere between the binary code and the ther-
mometer code have also been proposed, e.g., linear and polynomial codes
[67, 68], but these are not covered in this thesis. One code, that also falls between
the binary and the thermometer code, and is commonly used in DAC design is the
N
segmented code. An N -bit binary code corresponds to a ( 2 – 1 ) -bit thermome-
ter code. Hence, the hardware complexity in the encoding circuitry increases
exponentially with N , and becomes excessively large for large N . An often used
86
Digital DAC building blocks
compromise is to use thermometer code for a few of the MSBs, say K of the
MSBs, and let the remaining N – K LSBs be binary weighted. This is referred to
as a segmented code. Assume a binary weighted input code
X b = [ b N – 1, b N – 2, …, b 0 ] , (5.20)
X s = [ s 2 K + N – K – 2, …, s N – K , s N – K – 1, …, s 0 ] , (5.21)
where
[ s N – K – 1, …, s 0 ] = [ b N – K – 1, …, b 0 ] , (5.22)
i for i ∈ [ 0, N – K – 1 ]
w s, i = 2 . (5.23)
2 N – K for i ∈ [ 2 + N – K – 2, N – 1 ]
K
t 12(b 1, b 0) = b 1 + b 0 ,
2
t 2(b 1, b 0) = b 1 , and
t 32(b 1, b 0) = b 1 ⋅ b 0 . (5.24)
87
Current-Steering DAC Implementations
binary-to-thermometer
encoder
latches,
binary code
switches,
and
MSBs current sources
delay
LSBs
(a)
latches,
encoders
switches,
and
current sources
delay
(b)
Figure 5.8 (a) Segmented and (b) multi-segmented current-steering DAC architecture.
88
Digital DAC building blocks
In Fig. 5.7(b) it is indicated that the truth table of the 2-3 encoder, with inputs b 0
and b 1 , appears twice in the truth table for the 3-7 encoder. The logic expression
describing the 3-7 encoder is
t 3j (b 2, b 1, b 0) = b 2 + t 2j (b 1, b 0) = b 2 ⋅ t 2j (b 1, b 0) ,
t 43(b 2, b 1, b 0) = b 2 = b 2 , and
t j + 4(b 2, b 1, b 0) = b 2 ⋅ t 2j (b 1, b 0) = b 2 + t 2j (b 1, b 0) , (5.25)
where j ∈ { 1, 2, 3 } .
In Fig. 5.7(c) it is indicated that the logic expressions for a 4-bit encoder can be
obtained in a similar fashion if a 3-bit encoder is available. For the general N -bit
encoder we have
t Nj (b N – 1, …, b 0) = b N – 1 ⋅ t Nj – 1(b N – 2, …, b 0) ,
t 2 N – 1(b N – 1, …, b 0) = b N – 1 , and
t Nj + 2 N – 1 (b N – 1, …, b 0) = b N – 1 + t Nj – 1(b N – 2, …, b 0) , (5.26)
t Nj (b N – 1, …, b 0) = b N – 1 + t Nj – 1(b N – 2, …, b 0) ,
t 2 N – 1(b N – 1, …, b 0) = b N – 1 , and
t Nj + 2 N – 1 (b N – 1, …, b 0) = b N – 1 ⋅ t Nj – 1(b N – 2, …, b 0) . (5.27)
89
Current-Steering DAC Implementations
b
0 t2
1
2
b t2
1
2
t3
(a)
b0 tN−1 tN b0 tN−1 tN
1 1
1 1
tN−1
N−1
bN−2 N−1
tN bN−2 t2N−1−1 tN
2 −1 N−1 N−1
2 −1 2 −1
tNN−1
2 +1
tNN−1
2 +1
tNN tNN
2 −1 2 −1
(b) (c)
Figure 5.9 (a) 2-bit encoder, (b) even order N-bit encoder, and (c) odd order N-bit encoder.
The boxes in (b) and (c) represents (N-1)-bit encoders.
90
DAC design strategies and measurement setup
1
Logic level
n n+1 n+2
Update instant
Figure 5.10 Nonoverlapping switch signals for controlling differential PMOS switches.
D Q+
Q−
D
Figure 5.11 Implementation of switch signal generator utilizing cross coupled NOR gates.
The second solution utilizes clocked differential cascode voltage switch (DCVS)
[73] latches, and is shown in Fig. 5.12. It is required that the inputs only change
their values while the clock signal, Φ , is high, i.e., when the clocked transistors
are turned off. This can be ensured by adding a latch at the input which is trans-
parent only during the high clock phase. By proper sizing of the transistors, the
crossover point between Q + and Q – can be chosen to avoid having both switch
transistors simultaneously shut off [21].
91
Current-Steering DAC Implementations
D D
F F
Q Q+
92
DAC design strategies and measurement setup
Figure 5.14 Tree clock network for driving the switch signal generators.
93
Current-Steering DAC Implementations
94
A 14-bit DAC in 0.35 µm CMOS
power clock
supply generator
oscilloscope
measurement DAC
board
spectrum
GPIB analyzer
PC
controller
(a)
clock single-ended RF
outputs transformer
digital input
DAC
load
resistors combined
output
digital supply analog supply
(b)
Figure 5.15 Setup for DAC measurements. The block diagram in (a) shows the PC con-
trolled measurement system, and a PCB used for DAC measurement is outlined
in (b).
are approximately 2.5 MHz and 6.5 MHz, respectively. A common property for
all spectra in Fig. 5.17 is that the dominating spurious tone is the third harmonic,
and we observe that the SFDR decreases with increasing signal frequency.
95
Current-Steering DAC Implementations
Figure 5.16 Chip photo of the 14-bit current-steering DAC implemented in a 0.35 µm
CMOS process.
Output spectra for half-scale single-tone measurements are shown Fig. 5.18
below. The signal and update frequency for each sub figure is the same as in the
corresponding sub figure in Fig. 5.17. Once again we find that the dominating
spurious tone is the third harmonic, and that the SFDR decreases with increasing
signal frequency.
The measurements were carried out for several signal frequencies, and the SFDR
was extracted from the spectra. SFDR is plotted as a function of signal frequency
in Fig. 5.19. Fig. 5.19(a) shows SFDR for the full-scale inputs, whereas
Fig. 5.19(b) shows SFDR for the half scale inputs.
These measurement results correspond well with the simulation results presented
in Sec. 3.4. This indicates that the models of nonlinear behavior due to parasitic
capacitance that was presented provide good descriptions of the behavior of cur-
96
A 14-bit PRDEM DAC in 0.35 µm CMOS
−10 −10
−20 −20
−30 −30
PSD [dBm/Hz]
PSD [dBm/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Frequency [Hz] 6
x 10 Frequency [Hz] 6
x 10
(a) (b)
Output Spectrum, f = 20 MHz. Output Spectrum, f = 20 MHz.
u u
0 0
−10 −10
−20 −20
−30 −30
PSD [dBm/Hz]
PSD [dBm/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
0 2 4 6 8 10 0 2 4 6 8 10
Frequency [Hz] x 10
6 Frequency [Hz] x 10
6
(c) (d)
Figure 5.17 Single-tone full-scale measurements on 14-bit DAC in 0.35 µm CMOS. For (a)
and (b) the update frequency is 5 MHz, and the signal frequencies are approxi-
mately 0.63 MHz and 1.64 MHz, respectively. For (c) and (d) the update fre-
quency is 20 MHz, and the signal frequencies are approximately 2.55 MHz and
6.55 MHz, respectively.
97
Current-Steering DAC Implementations
−10 −10
−20 −20
−30 −30
PSD [dBm/Hz]
PSD [dBm/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Frequency [Hz] 6
x 10 Frequency [Hz] 6
x 10
(a) (b)
Output Spectrum, f = 20 MHz. Output Spectrum, f = 20 MHz.
u u
0 0
−10 −10
−20 −20
−30 −30
PSD [dBm/Hz]
PSD [dBm/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
0 2 4 6 8 10 0 2 4 6 8 10
Frequency [Hz] x 10
6 Frequency [Hz] x 10
6
(c) (d)
Figure 5.18 Single-tone, half-scale measurements on a 14-bit DAC in 0.35 µm CMOS. For
(a) and (b) the update frequency is 5 MHz, and the signal frequencies are
approximately 0.63 MHz and 1.64 MHz, respectively. For (c) and (d) the update
frequency is 20 MHz, and the signal frequencies are approximately 2.55 MHz
and 6.55 MHz, respectively.
5.6.1 Implementation
The PRDEM chip is constructed with four layers of switching. The control sig-
nals for the switches are generated off-chip to allow flexibility. The switching
tree is terminated with DAC banks consisting of one 1-bit DAC and one 10-bit
DAC, as was discussed in Sec. 4.1.3. The current sources and the current
switches are both of PMOS type, and no cascodes have been used in the current
98
A 14-bit PRDEM DAC in 0.35 µm CMOS
60 60
SFDR [dB]
SFDR [dB]
55 55
50 50
45 45
40 5 6 7
40 5 6 7
10 10 10 10 10 10
Signal frequency [Hz] Signal frequency [Hz]
(a) (b)
Figure 5.19 SFDR vs. signal frequency for (a) full-scale signals and (b) half-scale signals.
The upper curve in both figures correspond to the update frequency 5 MHz,
whereas the lower curves correspond to the update frequency 20 MHz.
sources. A description of the design strategy used is given in [76]. A chip photo
of the PRDEM DAC is shown in Fig. 5.16. The core area of the chip is approxi-
mately 6.3 mm2.
DAC bank
99
Current-Steering DAC Implementations
100
A 14-bit PRDEM DAC in 0.35 µm CMOS
−10
SFDR = 59.6 dB
−20
−30
PSD [dB/Hz]
−40
−50
−60
−70
−80
−90
−100
0 1 2 3 4 5
Frequency [MHz]
(a)
Simulated With 1 Layer DEM Simulated With 4 Layer DEM
0 0
−10 −10
SFDR = 67.8 dB SFDR = 68.1 dB
−20 −20
−30 −30
PSD [dB/Hz]
PSD [dB/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
−100 −100
0 1 2 3 4 5 0 1 2 3 4 5
Frequency [MHz] Frequency [MHz]
(b) (c)
Figure 5.21 Simulated spectra using 10MHz update frequency, (a) without randomization,
with (b) one switching layer and (c) four layer switching.
Measurement results
To verify the simulated results, the implemented DAC is measured with the same
update frequency, f u = 10 MHz . In Fig. 5.22(a) and Fig. 5.22(b) we compare
the gain in performance between using no randomization and using one switch-
ing layer. We find that the SFDR is increased from 60.7 to 67.7 dB, an improve-
ment of 7.5 dB which could be predicted from the simulation results. When
switching all four layers (Fig. 5.22(c)) we do not gain much in SFDR perfor-
mance compared with the results in Fig. 5.22(b).
101
Current-Steering DAC Implementations
−10
SFDR = 60.7 dB
−20
−30
PSD [dB/Hz]
−40
−50
−60
−70
−80
−90
−100
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Frequency [MHz]
(a)
Measured With 1 Layer DEM Measured With 4 Layer DEM
0 0
−10 −10
SFDR = 67.3 dB SFDR = 68 dB
−20 −20
−30 −30
PSD [dB/Hz]
PSD [dB/Hz]
−40 −40
−50 −50
−60 −60
−70 −70
−80 −80
−90 −90
−100 −100
0.5 1 1.5 2 2.5 3 3.5 4 4.5 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Frequency [MHz] Frequency [MHz]
(b) (c)
Figure 5.22 Measured spectra, 10MHz update frequency, (a) without randomization, with
(b) one switching layer and (c) four layer switching.
SFDR without randomization are compared in Fig. 5.23(a). The same compari-
son with randomization is shown in Fig. 5.23(b). We find from Fig. 5.23(a) and
Fig. 5.23(b) that the simulated and measured results behaves similarly, which is a
good validation of the models used in the simulations.
102
A dual 14-bit DAC in 0.25 µm CMOS
Simulated and measured SFDR without DEM Simulated and measured SFDR with DEM
65
Measured 74 Measured 12
Simulated Simulated
71
62 10
68 11
Number of bits
Number of bits
SFDR [dB]
SFDR [dB]
65
59
62 10
56 9 59
56 9
53
53
1 5 10 30 1 5 10 30
Sampling Frequency [MHz] Sampling Frequency [MHz]
(a) (b)
Figure 5.23 Simulated and measured SFDR performance for different update frequencies
with (a) no randomization, and (b) switching in all layers.
103
Current-Steering DAC Implementations
nal frequency in Fig. 5.25. For the cases when the update frequency is 5 and
10 MHz, SFDR is higher than 73 dB for a range of signal frequencies up to
3 MHz. This can be regarded as a good result, even if an SFDR of at least 86 dB
is required for 14 effective bits. The linearity is deteriorated when the update fre-
quency is increased to 25 MHz. This phenomenon is not yet explained, and
requires more study.
A more adequate measure for DMT applications is the MTPR, which is defined
in Sec. 2.3.2. For this purpose we use an ADSL-like input, which is quantized
values of
255
j⋅∆
X ( n ) = X DC + K ⋅ ∑ a j ⋅ sin 2πn ------------
fu
-f + φ j ,
(5.28)
j=1
104
Summary of implemented DACs
70
SFDR [dB]
65
60
55
50 6 7
10 10
Signal frequency [Hz]
105
Current-Steering DAC Implementations
missing tone
106
6 Conclusions
The thesis covered selected topics in the areas modeling, error correction tech-
niques and implementation of CMOS DACs. The main focus was on the linearity
properties of current-steering DACs for telecommunication applications.
The application of DACs in DSL environments was outlined. The requirements
on high linearity in the DAC was illustrated by examining the influence of non-
linear distortion on QAM constellations. The ideal DAC was presented, as well
as some performance measures commonly used for characterizing communica-
tion DACs.
Different sources of nonlinear behavior in current-steering DACs were discussed.
The static errors dominate the low-frequency behavior, whereas the high-fre-
quency behavior is dominated by dynamic errors. Mismatch between circuit ele-
ments is one of the main sources of static nonlinear behavior. An analysis of the
influence of linearly graded matching errors on the performance of current-steer-
ing DACs with a certain layout strategy was made. This layout strategy is known
to have poor properties when it comes to suppressing matching errors, but was
useful in the example. The source of nonlinear behavior that was focused on in
this work was the parasitic capacitance in the transistors and the wires. The phe-
nomenon was first modeled with a state-space model in Matlab. Results from
simulations using this model were presented to illustrate the effect of the noni-
deal components included in the model. A model with relatively low computa-
tional complexity was developed. This model is suitable for on-chip
implementation for use in error correction/compensation techniques.
Various methods for improving the linearity of DACs were discussed. Some of
the methods, like DEM and the differential DAC architecture, rely on redundant
coding to improve the linearity. In DEM the distortion caused by mismatch is, in
some sense, transformed into noise. To gain in the overall SNDR, oversampling
107
Conclusions
and filtering of out-of-band noise are required. Other methods, like distributed
biasing and a proposed calibration technique, aim at reducing the mismatch
between the output currents from the current sources. Two ideas utilizing models
of the dynamic nonlinearity caused by nonzero output capacitance in the current
sources were proposed. Both ideas aimed for enhanced linearity at higher signal
frequencies. The first method utilized a feedback loop similar to delta-sigma
modulation, and was denoted distortion shaping. The second method was a type
of predistortion. The simulations showed promising results, and an initial mea-
surement of the distortion shaping method indicates that the models used are
good enough to be practically useful. More measurements, and possibly model
modifications, are, however, required to draw any further conclusions.
Three different implementations of current-steering DACs in CMOS were pre-
sented. The required building blocks were discussed separately, and overall
design considerations were presented. The first two DACs were fabricated in a
0.35 µm CMOS process. One of these DACs utilizes the PRDEM technique for
enhanced static linearity. The third converter was fabricated in a 0.25 µm CMOS
process. Distributed biasing of current sources was used in this implementation
to reduce the matching errors between the current sources. Knowledge from the
behavioral level modeling was utilized to improve the performance from one cir-
cuit implementation to the next. Measurement results show close resemblance
with the simulation results using the behavioral level models presented in this
thesis. The simulations and measurements on the PRDEM DAC are especially
interesting, since they illustrate the limited efficiency of DEM when dynamic
errors dominate, besides validating the behavioral level modeling.
108
References
[1] T. Starr, J.M. Cioffi, and P.J. Silverman, Understanding Digital
Subscriber Line Technology, Prentice Hall PTR, Upper Saddle River, NJ,
USA, 1999, ISBN 0-13-780545-4.
[2] W.Y. Chen, DSL Simulation Techniques and Standards Development for
Digital Subscriber Line Systems, Macmillan Technical Publishing,
Indianapolis, IN, USA, 1998, ISBN 1-57870-017-5.
[3] D.J. Rauschmayer, ADSL/VDSL Principles, Macmillan Technical
Publishing, Indianapolis, IN, USA, 1999, ISBN 1-57870-015-9.
[4] S. Haykin, Communication Systems, 4th Ed., John Wiley & Sons, New
York, NY, USA, 2001, ISBN 0-471-17869-1.
[5] A. Bindra, “High-speed wideband DACs permit multicarrier cellular
basestations,” Electronic Design, Dec. 18, 2000.
[6] L. Wanhammar, DSP Integrated Circuits, Academic Press, San Diego, CA,
USA, 1999, ISBN 0-12-734530-2.
[7] M. Gustavsson, J.J. Wikner, and N. Tan, CMOS Dataconverters for
Communications, Kluwer Academic Publishers, Boston, MA, USA, 2000,
ISBN 0-7923-7780-X.
[8] D.A. Johns and K. Martin, Analog Integrated Circuit Design, John Wiley
& Sons, New York, NY, USA, 1997, ISBN 0-471-14448-7.
[9] R.J. van de Plassche, Integrated Analog-to-Digital and Digital-to-Analog
Converters, Kluwer Academic Publishers, Boston, MA, USA, 1994,
ISBN 0-7923-9436-4.
109
[10] P. Hendricks, “Specifying communication DACs,” IEEE Spectr., vol. 34,
no. 7, pp. 58-69, Jul. 1997.
[11] Defining and Testing Dynamic Parameters in High-Speed ADCs, Part 1,
Maxim Integrated Products, Inc. [Online]. Available: http://dbserv.maxim-
ic.com/tarticle/view_article.cfm?article_id=111.
[12] M. Schetzen, The Volterra and Wiener Theories of Nonlinear Systems,
Krieger Publishing Company, Malabar, FL, USA, 1989, ISBN 0-89464-
356-8.
[13] P. Wambaq and W. Sansen, Distortion Analysis of Analog Integrated
Circuits, Kluwer Academic Publishers, Boston, MA, USA, 1998, ISBN 0-
7923-8186-6.
[14] B.G. Streetman, Solid State Electronic Devices, Prentice-Hall, Englewood
Cliffs, NJ, USA, 1995, ISBN 0-13-436379-5.
[15] M.J.M. Pelgrom, A.C.J. Duinmaijer, and A.P.G. Welbers, “Matching
properties of MOS transistors,” IEEE J. Solid-State Circuits, vol. 24, no. 5,
pp. 1433-1439, Oct. 1989.
[16] A. Van den Bosch, M. Steyaert, and W. Sansen, “An accurate statistical
yield model for CMOS current-steering D/A converters,” Proc IEEE
International Symposium on Circuits and Systems (ISCAS’00), vol. 4,
pp. 105-108, Geneva, Switzerland, May 28-30, 2000.
[17] M.J.M. Pelgrom, H.P. Tuinhout, and M. Vertregt, “Transistor matching in
analog CMOS applications,” Proc. 1998 International Electron Devices
Meeting, pp. 915-918, 1996.
[18] A. Van den Bosch, M. Steyaert, and W. Sansen, “The extraction of
transistor mismatch parameters: the CMOS current-steering D/A converter
as a test structure,” Proc IEEE International Symposium on Circuits and
Systems (ISCAS’00), vol. 2, pp. 745-748, Geneva, Switzerland, May 28-30,
2000.
[19] H. Tuinhout, M. Pelgrom, R. Penning de Vries, and M. Vertregt, “Effects
of metal coverage on MOSFET matching,” Proc. 1996 International
Electron Devices Meeting, pp. 735-738, 1996.
[20] K.O. Andersson and J.J. Wikner, “Modeling of the influence of graded
element matching errors in CMOS current-steering DACs,” Proc.
NORCHIP Conference, pp. 399-404, Oslo, Norway, Nov. 8-9, 1999.
110
[21] G.A.M. Van der Plas, J. Vandenbussche, W. Sansen, M.S.J. Steyaert, and
G. Gielen, “A 14-bit intrinsic accuracy Q2 random walk CMOS DAC,”
IEEE J. Solid-State Circuits, vol. 34, no. 12, pp. 1708-1718, Dec. 1999.
[22] C.A.A. Bastiaansen, D.W.J. Groeneveld, H.J. Schouwenaars, H.A.H.
Termeer, “A 10-b 40-MHz 0.8-µm CMOS current-output D/A converter,”
IEEE J. Solid-State Circuits, vol. 26, no. 7, pp. 917-921, Jul. 1991.
[23] Y. Nakamura, T. Miki, A. Maeda, H. Kondoh, and N. Yazawa, “A 10-b
70-MS/s CMOS D/A converter,” IEEE J. Solid-State Circuits, vol. 26,
no. 4, pp. 637-642, Apr. 1991.
[24] I. Galton, “Spectral shaping of circuit errors in digital-to-analog
converters,” IEEE Trans. Circuits Syst. II, vol. 44, no. 10, pp. 808-817,
Oct. 1997.
[25] J.J. Wikner and N. Tan, “Modeling of CMOS digital-to-analog converters
for telecommunication,” IEEE Trans. Circuits Syst. II, vol. 46, no. 5,
pp. 489-499, May 1999.
[26] J.J. Wikner, CMOS Digital-to-Analog Converters for Telecommunication
Applications, Linköping studies in science and technology, Thesis No. 715,
ISBN 91-7219-277-1, Linköping, Aug., 1998.
[27] A. Van den Bosch, M. Steyaert, and W. Sansen, “SFDR-bandwidth
limitations for high speed high resolution current steering CMOS D/A
converters,” Proc. IEEE International Conference on Electronics, Circuits,
and Systems (ICECS’99), pp. 1193-1196, 1999.
[28] K.O. Andersson and J.J. Wikner, “Characterization of a CMOS current-
steering DAC using state-space models,” Proc. 43rd Midwest Symposium
on Circuits and Systems (MWSCAS’00), pp. 668-671, Lansing, MI, USA,
Aug. 8-11, 2000.
[29] H.B. Bakoglu, Circuits, Interconnections, and Packaging for VLSI,
Addison-Wesley, USA, 1990, ISBN 0-201-06008-6.
[30] L. Ljung, System Identification, Theory for the User, Prentice-Hall, Upper
Sadle River, NJ, USA, 1999, ISBN 0-13-656695-2.
[31] Control Systems Toolbox for use with Matlab - Reference Version 5,
Mathworks Inc., 2001, [Online]. Available: http://www.mathworks.com/
access/helpdesk/help/toolbox/control/control.shtml.
[32] A. Bugeja and B.S. Song, “A self-trimming 14-b 100-MS/s CMOS DAC,”
IEEE J. Solid-State Circuits, vol. 34, no. 12, pp. 1841-1852, Dec. 1999.
111
[33] AD9764, Datasheet, Analog Devices, Norwood, MA, USA, 1999.
[34] N.U. Andersson, K.O. Andersson, J.J. Wikner, and M. Vesterbacka,
“Models and implementation of a dynamic element matching DAC,” to
appear in Kluwer International Journal of Analog Integrated Circuits and
Signal Processing, Special Issue: Selected Papers from the NORCHIP
2001 Conference.
[35] N.U. Andersson, K.O. Andersson, J.J. Wikner, and M. Vesterbacka,
“Models and implementation of a dynamic element matching DAC,” Proc.
NORCHIP Conference, pp. 155-160, Kista, Sweden, Nov. 12-13, 2001.
[36] K.O. Andersson, “Improved current-steering D/A conversion,” Swedish
patent 0000731-0, 2000.
[37] D.W.J. Groeneveld and H.J. Schouwenaars, “A dual 3.4V bitstream
continuous calibration CMOS D/A converter with 110 dB dynamic range,”
Proc. 2nd IEE International Conference on Advanced A/D and D/A
Conversion Techniques and their Applications, pp. 42-47, 1994.
[38] C.-H. Lin and K. Bult, “A 10-b, 500-MSample/s CMOS DAC in 0.6 mm2,”
IEEE J. Solid-State Circuits, vol. 33, no. 12, pp. 1948-58, Dec. 1998.
[39] M.K. Rudberg, M. Vesterbacka, N. Andersson, J.J. Wikner, “Glitch
Minimization and Dynamic Element Matching in D/A Converters”, Proc.
IEEE International Conference on Electronics, Circuits, and Systems
(ICECS’01), Beirut, Lebanon, Dec. 17-20, 2000.
[40] M. Vesterbacka, M.K. Rudberg, J.J. Wikner, and N.U. Andersson,
“Dynamic Element Matching in D/A Converters with Restricted
Scrambling,” Proc. IEEE International Conference on Electronics,
Circuits, and Systems (ICECS’01), Beirut, Lebanon, Dec. 17-20, 2000.
[41] J.W. Bruce and P. Stubberud, “A comparison of hardware efficient
dynamic element matching networks for digital to analog converters,”
Proc. 43rd Midwest Symposium on Circuits and Systems (MWSCAS’00),
pp. 672-675, Lansing, MI, USA, Aug. 8-11, 2000.
[42] H.T. Jensen and I. Galton, “An analysis of the partial randomization
dynamic element matching technique,” IEEE Trans. Circuits Syst. II,
vol. 45, no. 12, pp. 1538-1549, Dec. 1998.
[43] P. Stubberud and J.W. Bruce, “An analysis of dynamic element matching
flash digital-to-analog converters,” IEEE Trans Circuits Syst. II, vol. 48,
no. 2, Feb. 2001.
112
[44] N.U. Andersson and J.J. Wikner, “A comparison of dynamic element
matching in DACs,” Proc. NORCHIP Conference, pp. 385-390, Oslo,
Norway, Nov. 8-9, 1999.
[45] J. Steensgaard, “High-resolution mismatch-shaping digital-to-analog
converters,” Proc. IEEE International Symposium on Circuits and Systems
(ISCAS’01), vol. 1, pp. 516-519, Sydney, Australia, May 6-9, 2001.
[46] U.K. Moon, G.C. Temes, and J. Steensgaard, “Digital techniques for
improving the accuracy of data converters,” IEEE Commun. Mag., vol. 37,
no. 10, pp. 136-143, Oct. 1999.
[47] J. Steensgaard, U.-K. Moon, and G.C. Temes, “Mismatch-shaping serial
digital-to-analog converter,” Proc IEEE International Symposium on
Circuits and Systems (ISCAS’99), vol. 2, pp. 5-8, Orlando, FL, USA, Jun.
1999.
[48] R.E. Radke, A. Eshraghi, and T.S. Fiez, “A 14-bit current-mode ∆Σ DAC
based upon rotated data weighted averaging,” IEEE J. Solid-State Circuits,
vol. 35, no. 8, pp. 1074-1084, Aug. 2000.
[49] A. Bugeja and B.S. Song, “A 14-b 100-MS/s CMOS DAC designed for
spectral performance,” IEEE J. Solid-State Circuits, vol. 35, no. 12,
pp. 1841-1852, Dec. 2000.
[50] M. Waltari, J. Pirkkalaniemi, M. Kosunen, L. Sumanen, and K. Halonen,
“A 14-bit, 40 MS/s DAC with current mode deglitcher,” Proc. NORCHIP
Conference, pp. 242-247, Kista, Sweden, Nov. 12-13, 2001.
[51] R.M. Gray and T.G. Stockham Jr., “Dithered quantizers,” IEEE Trans.
Inform. Theory, vol. 39, no. 3, pp. 805-812, May 1993.
[52] B. Brannon, Overcoming Converter Nonlinearities with Dither,
Application Note AN-410, Analog Devices, Norwood, MA, USA.
[53] K.O. Andersson, N.U. Andersson, M. Vesterbacka, and J.J. Wikner, “A
differential DAC architecture with variable common-mode level,” Proc.
IEEE International Symposium on Circuits and Systems (ISCAS’02),
vol. 1, pp. 113-116,Scottsdale, AZ, USA, May 26-29, 2002.
[54] K.O. Andersson, N.U. Andersson, M. Vesterbacka, and J.J. Wikner,
“Combining DACs for improved performance,” Proc. 4th IEE
International Conference on Advanced A/D and D/A Conversion
Techniques and their Applications, Prague, Czech Republic, June 26-28,
2002.
113
[55] K.O. Andersson, “Current Steering DAC,” pending Swedish patent, 2001.
[56] K.O. Andersson, N.U. Andersson, and J.J. Wikner, “Spectral shaping of
DAC nonlinearity errors through modulation of expected errors,” Proc.
IEEE International Symposium on Circuits and Systems (ISCAS’01),
vol. 3, pp. 417-420, Sydney, Australia, May 6-9, 2001.
[57] D. Anastassiou, “Error diffusion coding for A/D conversion,” IEEE Trans.
Circuits Syst., vol. 36, no. 9, pp. 1175-1186, Sep. 1989.
[58] R. Schreier and Y. Yang, “Stability tests for single-bit sigma-delta
modulators with second-order FIR noise transfer functions,” Proc. IEEE
International Symposium on Circuits and Systems (ISCAS’92), vol. 3,
pp. 1316-1319, 1992.
[59] K. Morris and P. Kennington, “Power amplifier linearisation using
predistortion techniques,” IEE Colloquium on RF and Microwave
Components for Communication Systems, Digest no. 1997/126, pp. 6/1-6/
6, 1997.
[60] S. Andreoli, H.G. McGLure, P. Banelli, and S. Cacopardi “Digital
linearizer for RF amplifiers,” IEEE Trans. Broadcast., vol. 43, no. 1,
pp. 12-19, Mar. 1997.
[61] H.W. Kang, Y.S. CHo, and D.H. Youn, “On compensating nonlinear
distortions of an OFDM system using an efficient adaptive predistorter,”
IEEE Trans. Commun., vol. 47, no. 4, pp. 522-526, Apr. 1999.
[62] K.J. Riley, D.M. Hummels, F.H. Irons, and A. Rundell, “Dynamic
compensation of digital to analog converters,” Proc. 16th IEEE
Instrumentation and Measurement Technology Conference (IMTC’99),
vol. 2, pp. 1310-1315, 1999.
[63] K.O. Andersson and J.J. Wikner, “Digital-to-analog converter having error
correction,” pending Swedish patent, 2002.
[64] J.M. Rabaey, Digital Integrated Circuits - a Design Perspective, Prentice-
Hall, Upper Saddle River, NJ, USA, 1996, ISBN 0-13-394271-6.
[65] N.H.E. Weste and K. Eshraghian, Principles of CMOS VLSI Design a
Systems Perspective, Addison-Wesley, USA, 1993, ISBN 0-201-53376-6.
[66] S.M. Kang and Y. Leblebici, CMOS Digital Integrated Circuits Analysis
and Design, McGraw-Hill, USA, 1996, ISBN 0-07-038046-5.
114
[67] J.J. Wikner and M. Vesterbacka, “D/A conversion with linear-coded
weights,” Proc. 2000 IEEE Southwest Symposium on Mixed-Signal Design
(SSMSD’00), pp. 61-66, San Diego, CA, USA, Feb. 28-29, 2000.
[68] M. Vesterbacka, K.O. Andersson, N.U. Andersson, and J.J. Wikner,
“Using different weights in DACs,” Proc. 4th IEE International
Conference on Advanced A/D and D/A Conversion Techniques and their
Applications, Prague, Czech Republic, June 26-28, 2002.
[69] J.J. Wikner, Studies on CMOS Digital-to-Analog Converters, Linköping
studies in science and technology, dissertation no. 667, ISBN 91-7219-
910-5, Linköping, Sweden, Apr. 2000.
[70] A. Van den Bosch, M. Borremans, M. Steyaert, and W. Sansen, “A 10-bit
1-GSample/s current-steering CMOS D/A converter,” IEEE J. Solid-State
Circuits, vol. 6, no. 3, pp. 315-324, Mar. 2001.
[71] K. Skahill, VHDL for Programmable Logic, Addison-Wesley, USA, 1996,
ISBN 0-201-89573-0.
[72] J. Bastos, A.M. Marques, M.S.J. Steyaert, W. Sansen, “A 12-bit intrinsic
accuracy high-speed CMOS DAC,” IEEE J. Solid-State Circuits, vol. 33,
no. 12, pp. 1959-1969, Dec. 1998.
[73] K. Bernstein, et. al., High Speed CMOS Design Styles, Kluwer Academic
Publishers, Kluwer Academic Publishers, Boston, MA, USA, 1999,
ISBN 0-7923-8220-X.
[74] X. Aragones, J.L. Gonzalez, and A. Rubio, Analysis and Solutions for
Switching Noise Coupling in Mixed-Signal ICs, Kluwer Academic
Publishers, Boston, MA, USA, 1999, ISBN 0-7923-8504-7.
[75] O. Andersson, Mismatch Modeling and Design of CMOS Current-Steering
Digital-to-Analog Converters, M.Sc. thesis, Linköping, Nov., 1999.
[76] N.U. Andersson and J.J. Wikner, “A Strategy of Implementing Dynamic
Element Matching in Current-Steering DACs,” Proc. 2000 IEEE
Southwest Symposium on Mixed-Signal Design (SSMSD’00), pp. 51-56,
San Diego, CA, USA, Feb. 28-29, 2000.
[77] J. Vandenbussche, G. Van der Plas, W. Daems, A. Van den Bosch,
G. Gielen, M. Steyaert, and W. Sansen, “Systematic design of high-
accuracy current-steering D/A converter macrocells for integrated VLSI
systems,” IEEE Trans. Circuits Syst II, vol. 48, no. 3, pp. 300-309, Mar.
2001.
115
[78] G.E.G. Gielen, “New methods and tool for analog IP use in mixed-signal
integrated systems,” Proc. NORCHIP Conference, pp. 5-9, Kista, Sweden,
Nov. 12-13, 2001.
116
Systematic Design of a 14-bit 150-MS/s
CMOS Current-Steering D/A Converter
G. Van der Plas, J. Vandenbussche, W. Daems, A. Van den Bosch,
*
G. Gielen , M. Steyaert, W. Sansen
Department of Electrical Engineering, Katholieke Universiteit Leuven, ESAT-MICAS
Kardinaal Mercierlaan 94, 3001 Heverlee, Belgium
E-mail : georges.gielen@esat.kuleuven.ac.be
Abstract II Current-steering D/A Converter Architecture
This paper presents a D/A converter with a 14-bit in- For high-speed, high-accuracy D/A converters, a segmented cur-
trinsic linearity in 0.5µm CMOS technology, which has rent-steering topology is usually chosen as it is intrinsically faster
been designed using a systematic design methodology and more linear than competing architectures [2]. The principle of
this type of D/A converter is depicted in Fig. 1: the l least signifi-
for current-steering D/A converters. A flexible architec- cant bits are binary implemented, the m most significant bits steer
ture is proposed for which the design parameters are a unary current source array.
calculated using a performance-driven top-down design VDD
t5
ILSB 2ILSB 4ILSB E IMSB IMSB IMSB IMSB
b8 t6
T
I Introduction b9 E t7
b10 R t8 t251
b3 b4 b5
In this paper the design methodology and the design of a complete b11
D
8ILSB 16ILSB 32ILSB b12 IMSB ... ... IMSB
current-steering D/A converter implemented in a standard 0.5µm b13
E
C t251
CMOS technology, is presented. This is the first publicly reported O
D
t252
t252 t253 t254
CMOS D/A converter with 14-bit intrinsic linearity, achieved binary weighted E
R
t253
t254 IMSB IMSB IMSB
without trimming or tuning [1]. The D/A converter has been de- current cells
signed according to a systematic design methodology which al-
unary current cells
lows to generate new designs for given specifications, or to easily
port designs to new processes. The design methodology covers Fig. 1: Simplified block diagram of a 14-bit segmented current-steering
the complete design flow and is supported by both commercial D/A converter. The l (=6) least significant bits steer the binary
and in-house developed academic software tools to speed up the weighted current sources directly. The m (=8) most significant bits
task significantly. Design times of less than one month total ac- are fed into the thermometer decoder, which steers the unary current
source array
cumulated person effort are obtained. The inputs are the specifica-
tions of the D/A converter and the technology and the result is a Specification Unit Value
verified layout. Static Number of bits (n) - 14
This paper is organized as follows. In section II the architecture of INL LSB 0.5
the D/A converter and its design parameters are described. Sec- DNL LSB 0.5
tion III explains the proposed systematic design methodology. Parametric Yield % 99.9
The sizing synthesis is explained to its full extent in section IV. Dynamic Sample frequency MHz 100
Section V describes the layout generation process. Section VI SFDR@500kHz dB 80
presents the experimental results and measurements of the im- Settling time ns 10
plemented design. Conclusions are given in section VII. Environmental Output range (Vswing) V 0.5
Permission to make digital/hardcopy of all or part of this work for personal or RLoad Ω 25
classroom use is granted without fee provided that copies are not made or
Digital levels - CMOS
distributed for profit or commercial advantage, the copyright notice, the title of the
publication and its date appear, and notice is given that copying is by permission of Power Supply V 2.7
ACM, Inc. To copy otherwise, to republish, to post on servers or to redistribute to Technology - 0.5µ 1P3M
lists, requires prior specific permission and/or a fee.
DAC 2000, Los Angeles, California
Optimization Power Consumption mW Minimize (300)
(c) 2000 ACM 1 -58113-188-7/00/0006..$5.00 Area mm2 Minimize (10)
Table 1: General specification list for a current-steering D/A converter with
typical values.
*
research associate of the National Fund of Scientific Research (Belgium)
The general specification list for a current-steering D/A converter The last important architectural design parameter of a current-
is given in Table 1. The specifications can be divided into four steering D/A converter is the switching scheme. The switching
categories: static, dynamic, environmental and optimization scheme has two components. A unary current source can be split
specifications. In the case of a D/A converter the static parameters in one or more units, as shown in Fig. 3. By splitting the unary
include static accuracy (i.e. number of bits), integral non- current source the spatial errors are averaged out, which is neces-
linearity (INL), differential non-linearity (DNL) and yield. The sary for high-accuracy applications [3,4]. The second parameter of
dynamic parameters include sample frequency, spurious free dy- the switching scheme is the switching sequence. In [5] it is shown
namic range (SFDR) and settling time. The environmental pa- that the remaining spatial errors are not accumulating when the
rameters include the power supply, the digital levels, the output current sources are switched on in an optimal order when the
load and the input/output range. The power consumption and area input code increases. The here proposed architecture differs from
need to be minimized for the specified technology. This specifica- previously used architectures in that the switching is fully flexible
tion list serves as input for the design process as will be explained and can be programmed when generating the layout to optimally
in the following sections. compensate for systematic errors which would otherwise deterio-
rate the targeted linearity.
In the next paragraphs the proposed D/A converter architecture
and its design parameters will be presented in detail. The designable parameters of the presented D/A converter are
summarized in Table 2. They have been classified in architecture
b13 b12 b11 b10 b9 b8 b7 b6 b5 b4 b1 b0
... level parameters and circuit level parameters
Clock Thermometer decoder Latency
driver equaliser Design parameters of the converter
VDDcur_src Full Decoder Architecture level l (number of LSBs)
un 1...255 bin 1...6
m (number of MSBs)
Iout Number of units (1/4/16 units) for unary current
Iout Iout Binary latch
Unary latches & switches source
& switches
Iout
switching sequence
2b M2a Swatch Array Circuit level WM1, LM1, (Vgs-Vt)M1
255 6
WM2, LM2
M1 Binary
Unary current source current src Latch transistor sizes
array array Clock driver sizes
VSScur_src
Current Source Array Table 2: The designable parameters of the presented D/A converter.
Fig. 2: The block diagram and floorplan of the proposed architecture.
III Systematic Design Methodology
The simplified block diagram of Fig. 1 has been implemented by
The performance-driven top-down design methodology has been
the proposed segmented architecture shown in Fig. 2. The current
accepted as the de facto standard for systematically designing
generated by the current sources (M1) are switched to one of the
analog building blocks in academia [6,7]. In [8] the design of
two differential output nodes of switch transistors M2a and M2b.
current-steering D/A converters has been automated following
These are driven by a latch. The full decoder comprises the ther-
this methodology for a specific architecture which is however
mometer decoder, which generates the steering signals for the
only feasible for 8 or 10-bit D/A converters.
unary latches, and a latency equalizer block. This latency equal-
izer block ensures correct timing for the steering signals of the
binary latches. One of the important architectural design parame- Analog Digital
ters is how many bits are implemented with binary weighted cur- Sizing
Synthesis
rent sources and how many with unary weighted current sources. Architectural level
The floorplan of the presented architecture is also shown in Fig. 2. Verification Verification
The switches and latches are implemented as one unit cell, and
placed in an array, referred to as the “swatch” array. The current
Sizing
source transistors are also placed in an array, the current source Device level
array. The three large modules (full decoder, swatch array and
current source array) are connected by signal busses. A clock Verification
driver completes the D/A converter.
Floorplanning
Verification Verification
80 Dynamic performance
70
In order not to deteriorate the dynamic performance, the following
60
factors are taken into account in the circuit level synthesis [12]:
Yield [%]
1 2 2
4 AVT
W∗ L = βA + 2 (3) Design parameters of the architecture
2σ 2 ( I ) (VGS − VT ) Architecture level 6 (number of LSBs)
I2 8 (number of MSBs)
16, Quad Quadrant [1] (number of units)
where σ(I)/I is the derived unit current source standard deviation
Random Walk [1] (INL optimized
and Aβ , AVT are technology constants [11]. For minimal area the switching sequence)
Vgs-Vt is maximized.
Circuit level WM1=1.1µ, LM1=104.5µ, (Vgs-Vt)Mcur_src=1Volt
The lower bound for the area is given by: WMsw=0.8µ, LMsw=1µ
sized latch transistor sizes
1
(W ∗ L) est ≈ Aβ2 (4) sized clock driver
2σ 2 ( I )
Table 3: The sized values of the designable parameters of the proposed D/A
I2 converter architecture.
The total current source array area is then estimated to be at least:
E. Sizing Result
areacur _ src ≈ 2 n (W ∗ L) est f routing (5) In Table 3 the final sizing results are summarized for the specifi-
cations of Table 1.
V Layout Generation wires. The connection pattern has to be entered in a C-like format,
and the tool automatically generates the actual physical layout.
Current-steering D/A converters are a typical example of layout-
driven analog design. The sized schematic alone does not consti- The current source unit is the basic cell in our case. It is placed in
tute an operational converter. An important part of the perform- a regular array and the parallel units are connected with the de-
ance is determined by the handling of layout-induced parasitics scribed bus routing. In total over 4000 units are placed and con-
and error components (i.e. systematic errors). All classical coun- nected with more than 1000 wires. Doing this task manually is
termeasures for digital to analog coupling (guard rings, shielding, impossible. The CPU time of Mondriaan on a standard SUN Ul-
separate supplies, …), and matching guidelines (equal orientation, tra I-166 workstation is less than 2 minutes (including the transfer
dummies, …) have been applied and will not be further discussed. to a commercial EDA framework). The result is a layout and a pin
list, which serves as input to the swatch array layout generation.
A. Floorplanning
The floorplan proposed in Fig. 2 is now refined. The relative Swatch Array Layout Generation
placement of the modules (full decoder, swatch array, current One swatch cell has been manually laid out. The placement and
source array) was already fixed and the actual sizes (area) of the routing of the swatch array is then done automatically with the
modules are readily available. However the aspect ratio is still to Mondriaan tool [13]. Inputs to the tool are the pin list of the cur-
be determined. First of all the global aspect ratio influences the rent source array, the netlist and the swatch cell layout. From this
aspect ratio of the modules. In general a square or near square the placement of the cells is derived and the routing created. The
chip layout is preferred. Secondly, at the chip-level the connec- output is the layout and a pin list of the bus for the steering signals
tions between the modules are extremely important: a fixed pitch coming from the decoder.
for the three modules must be chosen to route the busses between
the different modules to avoid wiring area loss. Furthermore, the Decoder Standard Cell Place and Route
choice of the chip-level pitch also ensures that the modules will
have the same width, resulting in an elegant chip assembly as can The layout of the digital decoder is generated using a standard cell
be seen in Fig. 9. place and route tool, i.e. the tools from Avant!. The pin list ob-
tained from the swatch array layout is input to the floorplan for
B. Circuit and Module Layout Generation the layout generation.
The current source cell array, swatch cell array and full decoder
are generated next. C. Layout Assembly
The modules are placed stacked on top of each other. The bus
Current Source Layout Generation generators of the Mondriaan tool [13] are used to generate the
actual connections between the three modules (full decoder,
The sizes of the unit current source have been determined. From swatch and current source array). Trees are used to collect the
this the sizes of all other weighted current sources and the unary output signals and distribute the clocking signal from the clock
current source are derived. To have optimal matching properties driver to the swatch array to have equal delay.
the current source must be built up from identical basic units. This
basic unit is manually laid out. Also the floorplan of the current
source array is determined (number of units, switching sequence).
contact
area
Y
X-Y
Fig. 8: Symbolic view of cells and routing channels: vertical (Y) routing
across cells, horizontal (X) routing in between cells. Vertical wires
connect to the contact areas in the cell, horizontal wires connect to
the vertical wires.
I. INTRODUCTION
range (SFDR) is one of the most important properties for these The variation of the output voltage of the DAC will result in
DACs. Very high linearity is required in order to achieve sat- the variation of the moment when the switches steering the cur-
isfying SFDR. For this purpose, the nonlinearities which may rent from the current sources to either of the two outputs change
impact the SFDR have to be analyzed and kept under control their status. We call this variation the ODDD in this paper. The
by the designers. physical reason of the ODDDs is as follows. Fig. 1 shows the
The current-steering DAC architecture is almost exclusively basic schematic of a current-source cell used in current-steering
used for high-speed high-accuracy DACs [1]. Two kinds of non- DACs. Normally the switch transistors work in saturation re-
linearities have been distinguished for this kind of DACs [2]. gion (when on) or cutoff region (when off). The voltage of the
One is the limited output impedance of the current sources, internal node X depends on the drain voltage of the on transistor.
which has been well analyzed and solved by taking the dif- Thus, the variation of the output voltage will cause a small vari-
ferential output [3]. The other kind is the delay-related nonlin- ation of , which will result in a change of the next switching
earities. According to the different causes of the delay differ- time.
ences, the cell-dependent delay differences (CDDDs) and the For example, assume that in the current sampling period,
output-dependent delay differences (ODDDs) are distinguished is high and is low, so the left transistor is on.
in our previous paper [2]. A mathematical method is presented Different values of will result in different values of .
in [2] to analyze the impact of the CDDDs on the SFDR of a cur- When the next sampling period comes, is getting lower
rent-steering DAC. Formulas with clear physical meaning were to turn off the left switch. The switch will not be turned off
derived, and were verified by comparing them to simulation re- until reaches the value of , where is
sults and published measurement results [4]. In this paper, the the threshold voltage of the switch transistors. On the other
method proposed in [2] is further extended to the case of the side, the right switch will not be turned on until reaches
ODDD. As in [2], formulas with clear physical meaning will be . Thus, the delay of the next sampling period will
obtained and verified by simulations, and then solutions to im- depend on the value of hence . This dependence
prove the SFDR will be discussed based on these formulas. will cause harmonic distortion on the output of the DAC and
therefore deteriorate the SFDR.
Manuscript received March 20, 2006; revised August 4, 2006. This paper was When in Fig. 1 is high, circuit analysis gives
recommended by Associate Editor T. N. Tarim.
The authors are with ESAT-MICAS, Katholieke Universiteit Leuven, Heverll
3001, Belgium (e-mail: tao.chen@ieee.org).
Digital Object Identifier 10.1109/TCSI.2006.887598
(1)
where is the transconductance and output resistance where Taylor series expansion has been applied. Higher order
of the switch transistor respectively, and is the output terms are omitted, since it should always be true for a good
impedance of the cascoded current source (see Fig. 1). The design that . This condition also means that, if a term
value of is normally in the order of 10 or lower for very in (3) with low order contributes to harmonic distortion at some
deep-submicrometer technologies. Assume for example the frequency, the impact of the higher order terms in (3) on that
variation of the output voltage is 0.5 V, and is 10, harmonic distortion will be very small and negligible.
then the variation of will be up to 0.05 V. If the rising
speed of the control signal is 1.25 V/ns, the 0.05 V A. SFDR of the DAC’s Single-Ended Output
variation of will result in a delay difference of up to
40 ps. This is already a very large value for a 14-bit DAC with As will be seen later, the first term of (3) will result in the
sampling frequency higher than 100 MHz. The impact on the second-order distortion, the second term in (3), which only
dynamic properties will be investigated in detail in this paper. slightly impacts the amplitude of the second-order distortion,
In Section II, a mathematical analysis is presented to analyze can be omitted
the impact of the ODDDs on the SFDR of both the single-ended
and differential outputs of a full-unary DAC. Then the behav-
ioral-level simulations (with and without glitches) are discussed
in Section III. Both the results of calculations and simulations (4)
are compared in this section. In Section IV, the design require- Assume the DAC’s input is a sinusoidal signal
ments imposed by this nonlinearity are given, and possible solu-
tions are summarized, some of which have already been applied
in publications from different authors. Finally, Section V draws (5)
conclusions.
where is the DAC’s number of bits. For convenience, the am-
plitude is set so that the amplitude of every unit current source
II. MATHEMATICAL ANALYSIS is 1. This will not impact the results since the SFDR is a ratio.
For a high-resolution DAC, ignoring the quantization errors,
In the following analysis, we assume an ideal full-unary cur- the ideal output current (which is an integer due to the assump-
rent-steering DAC, whose only nonlinearity is the ODDDs.1 We tion of (5)) of the DAC in the period from to
compare this to other effects in the conclusions. can be expressed as
Consider one of the current sources . Assume it is switched
on in the sampling cycle with delay . Then its output
current is (6)
(7)
(2)
(8)
where is the sampling period, is the time constant at the (9)
output node determined by the DAC’s load. The amplitude is
set to be one. This is a simplified expression, the internal poles During this period the full-unary DAC’s total second-order
of the current cell have been omitted, but it’s sufficient to get distortion in the single-ended output currents is contained in the
meaningful results. equation below
The distortion of this current source due to the delay is thus
(10)
(3) when
(11)
1For a full-binary or segmented DAC, the dynamic property should be
else
worse than a full-unary DAC due to the extra nonlinearity caused by the
binary-weighted current sources. is a square function.
270 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 2, FEBRUARY 2007
The results [shown in Fig. 4(d), later] show that only slightly
impacts the SFDR of a DAC without glitches.2
(12)
(13)
(14)
b) When , the value of depends on , the
delays of the two switches can be expressed as
(15)
(16)
For the case (a), applying (13) into (10) gives
(17)
Fig. 2(a) shows the waveform of the control signals and
during the transition. The moment when the switches change For the case (b), applying (15) into (10) gives
their status depends on their common-source voltage, i.e., in
Fig. 1. The value of is determined by the voltage of the output
end whose switch is on. In the case of Fig. 2(a), is determined
by since is high before the transition and therefore the
corresponding switch is on. and are the minimum (18)
and maximum values of , corresponding to the minimum and
maximum values of respectively. If we select the moment With the condition of (12), (18) can be simplified into (17).
when the off switch is turned on at the maximum (that is The second term in (17) only slightly impacts the amplitude at
) to be time 0 as a reference, then, according to (1), when the signal frequency and can be omitted. Thus, the distortion can
, this delay can be expressed as where is a be expressed as
proportionality factor. For the other switch, which is turning off,
the delay is between and , where is the delay
difference between the two transitions (on to off and off to on).
can be adjusted (by the designer) by changing the crossing
point of the two control signals and . If is adjusted to
the value , there will be no delay difference between the
two transitions. This is used by some designs [5], [4] to avoid the
situation that both the switches are on or both are off, thus the
current of the current sources can be kept stable. For simplicity, (19)
we adopt this condition in our analysis as shown in (12) and
2In an actual current-source cell, the different delays between the on-to-off
Fig. 2(b), meaning that by design the switches are on/off at the
same moment and the current is constant. The impact of on
and off-to-on transitions (D 6 0d
= ) will result in the variation of the total
current, which will result in glitches and deteriorate the dynamic properties of
the SFDR will be investigated by simulations in Section III-A. the DAC.
CHEN AND GIELEN: ANALYSIS AND IMPROVEMENT OF CURRENT-STEERING DAC’S DYNAMIC SFDR—II 271
The second term in the bracket of (19) only affects the dc level
and can be omitted. So (27)
(29)
(31)
The impact on the differential-output SFDR of the two terms in
(3) and (31) will be analyzed separately, assuming the condition
(25) of (12).
1) Impact of the First Terms: We will firstly analyze the first
where is the clock frequency. terms of (3) and (31). They are shown below
The Fourier transform of (22) is given by
(32)
(33)
where the subscription “ ” is used to clarify that this distortion
(26)
comes from the first terms of (3) and (31).
272 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 2, FEBRUARY 2007
With the output current expressed in (6)–(9), the total distor- Adding up the distortions in all the sampling cycles, the total
tion generated by (32) and (33) during this sampling cycle is distortion caused by the first terms can be obtained as
(34)
(39)
Applying the Fourier transform to (39) gives
(35)
where is defined in (11).
The total distortion in the th sampling cycle of the differen-
(40)
tial output is then
It is clearly shown in (40) that no harmonic can be gener-
ated from the first terms of (3) and (31). That means that we
have to take the higher order harmonics (the second terms) into
consideration.
2) Impact of the Second Terms: The second terms of (3) and
(31) are
(41)
(42)
(36)
where the subscription “ ” is used to clarify that this distortion
comes from the second terms of (3) and (31).
For the case (a) defined in (13) and (14)
The current output of the DAC is expressed in (6)–(9). During
the n’th sampling cycle the DAC’s distortion generated by (41)
and (42) in both the two output currents is
(37) (43)
where the condition of (12) has been applied, and is de-
fined in (20).
With (6)–(9), (37) can be simplified into
(38)
(44)
For the case (b) defined in (15) and (16), similar calculations
give the same result as (38). So this equation describes the dis- where (41), (42) and (11) have been used.
tortion (during the n’th sampling cycle) caused by the first terms Similar to the previous analysis, there are also two cases here:
of (3) and (31) in the DAC’s differential output. (a) ; (b) .
CHEN AND GIELEN: ANALYSIS AND IMPROVEMENT OF CURRENT-STEERING DAC’S DYNAMIC SFDR—II 273
For the case (a), applying (43), (44), and (13)–(16), the third- From (50), the distortion for all cycles can be obtained
order distortion of the differential output is
(45)
(51)
For the case (b), with similar calculations, the third-order dis-
tortion is
(52)
With the results of (23), (25), and (52), the Fourier transform
of (51) can be obtained
(47)
Applying (12) into (46) and (47), the same result is obtained
(48) (53)
Similar analysis as in Section II-B1) shows that the second This is a sequence of Dirac functions at the frequencies
term in the braces of (48) does not generate any harmonic dis- . What we are interested in is the component at the frequency
tortion. Only the first term needs to be considered . Its amplitude is
(54)
(50) (56)
274 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 2, FEBRUARY 2007
(57)
(58)
Fig. 4. Single-pole-model simulations: the deterioration of the SFDR caused by ODDDs as a function of the different parameters. (a) SFDR versus signal fre-
quency: d 0 0
= 40 ps, D = d ; = 0:25 ns. (b) SFDR versus maximum delay difference: f = 21:5 MHz, D = d ; = 0:25 ns, (c) SFDR versus
output time constant: f = 21:5 MHz, d 0
= 40 ps, D = d . (d) SFDR versus on-to-off and off-to-on delay difference: f = 21:5 MHz, d = 40 ps,
= 0:25 ns.
A behavioral model of DACs was presented in [8], where the way that ODDD impact the SFDR since the DAC’s output
the glitches are included as an exponentially damped sine and a during transition is changed by the glitches. In the model of
shifted hyperbolic tangent (59), the glitch period describes the time needed for the DAC
to transit from one output value to another; it can be extracted
from the results of SPICE simulations.
In our simulations the time-domain output of each transition
(59) is obtained by (59), after which it is sampled to do the FFT. The
sampling points are carefully selected so that one of the samples
in which is the output current, and is the amplitude is located in the middle of the transition. This is to achieve the
and period of the glitch, respectively. and are the highest accuracy, since the value in the middle of the transition
code levels between which the converter switches. is the delay is very sensitive to the delay of the transition.
of the transition. Repeating this process for each cycle, we can get the whole
Our simulations are based on this model. In our simulations, time-domain output of the DAC. A Hanning window is used be-
is set as described in Section III-A to include the ODDDs. fore the FFT analysis to reduce the time length needed while
For full-unary DACs can be set to be proportional to the maintaining enough accuracy. From the power spectrum ob-
variation of the codes, i.e., to . We define the tained, the SFDR at that signal frequency can be obtained.
proportionality factor pAgl as below The impact of the glitches is shown in Fig. 5. Fig. 5(a) shows
the impact of the glitches on the SFDR of a full-unary DAC
without ODDDs . As we expect, the glitches do
pAgl (60) not impact the SFDR for both the single-ended and differential
output.
So the glitches themselves are a kind of linear distortion and In Fig. 5(b), the curve with star markers is the case of the
will not impact the SFDR. However, the glitch will change single-ended output. The glitch has little impact on its SFDR
276 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 2, FEBRUARY 2007
since the ODDD itself has already deteriorated the SFDR while for the differential output this slope is about dB/dec.
greatly (see Fig. 4). The curve with circle markers shows the These slopes are consistent with our mathematical results ((30)
impact of the glitches on the SFDR of the differential output of and (57)), although the mathematical calculations are based on
DACs with ODDD. The SFDR reduces quickly with increasing the simplified first-order model.
amplitude (energy) of the glitches, hence illustrating the impact Two conclusions can be drawn from these results as follows.
of glitches. • Even for DACs with glitches, the impact of ODDD on
The SFDR- and SFDR- curves with different values of SFDR still obeys similar rules as predicted by our math-
pAgl are shown in Fig. 6. In these simulations was chosen ematical model ((29) and (56)), i.e., the slope of the falloff
to be 0.41 ns so that the simulated SFDR when pAgl (that remains the same (but the magnitude is different).
means no glitches exist in the output) is the same as the SFDR • The glitches can deteriorate the impact of ODDD on the
calculated when ns. In this way the simulation re- SFDR seriously, especially in the case of the differential
sults in this section can be easily compared with the results output.
in Sections II and III-A, where the analysis was based on the All these simulations in Figs. 5 and 6 are for full-unary DACs.
first-order model. The latter calculation results are also drawn For segmented DACs with binary LSBs, the glitches themselves
in Fig. 6 for comparison reasons. become a kind of nonlinear distortion and the SFDR can be
Fig. 6(a) shows the SFDR- curves with different pAgl expected to be even worse.
values. The SFDR decreases with increasing signal frequency
at a slope of about dB/dec, which is the same as the results IV. APPLICATION OF THE RESULTS
of the mathematical model based on the first-order model. In order to get some quantitative idea about the maximum
For the differential output, the SFDR is seriously reduced permitted , we will present some calculations with actual
when the amplitude of the glitches increases. The SFDR- values. For example, assume we want to design a 14-bit DAC
curves with different pAgl values are shown in Fig. 6(b). For with signal frequency up to 75 MHz and SFDR higher than
the single-ended output, the SFDR decreases with increasing dB. Assume the time constant at the
maximum delay differences at a slope of about dB/dec, output node is ns, then with (56) the maximum
CHEN AND GIELEN: ANALYSIS AND IMPROVEMENT OF CURRENT-STEERING DAC’S DYNAMIC SFDR—II 277
(63)
(66)
(62)
According to (66), we can reduce the maximum ODDDs by
This points us to some solutions to improve the SFDR lim- reducing the sizes of the switch transistor and increasing its gate
ited by ODDD. In (62) the factor is the switching overdrive voltage (to reduce the value of ). For today’s de-
speed, so the first solution is to accelerate the switching speed. signs, the latter is difficult to achieve since the available voltage
The intrinsic idea of this method is to reduce in (62). This is very limited. But the smallest sizes should be used for the
solution is for instance used in [9] by applying a David-Goliath switches (to reduce ).
latch. Equation (56) also shows that increasing the output time con-
The second possible solution is using a return-to-zero (RZ) stant (by increasing the load capacitance) can also reduce the
output stage as presented in [10]. With the RZ output stage, harmonic distortion caused by the ODDDs. But this reduces the
the glitches are greatly reduced. Furthermore, folding current overall bandwidth.
sources with gain-boosting are applied in this output stage to Finally, enough attention should be paid to the control of
reduce the variation of the output voltages at the drain nodes the glitch energy, even for DACs with their most MSBs imple-
of the switch transistors. So in (62) is reduced. This RZ mented in a unary way. This has been done in many published
output stage improves the dynamic property greatly, as shown designs. In [5] and [4], delay-controllable latches are used to re-
in [10] and [11]. The price is a lot of extra power consumption duce the glitches. The crossover adjustment circuit presented in
by the folded current sources and a significant reduction of the 3In [5] the cascoded switches are actually used for reducing the glitches in-
signal power due to the RZ operation. stead of increasing the gain of the switches.
278 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 54, NO. 2, FEBRUARY 2007
Tao Chen (S’02–M’07) was born in Zhangping City, Georges G. E. Gielen (S’87–M’92–SM’99–F’02)
Fujian Province, China, in 1974. He received the received the M.Sc. and Ph.D. degrees in electrical en-
B.Sc. degree in electronic engineering from Zhejiang gineering from the Katholieke Universiteit Leuven,
University, Hangzhou, China, the M.Sc. degree in Leuven, Belgium, in 1986 and 1990, respectively.
electronic engineering from Tsinghua University, In 1990, he was appointed as a Postdoctoral Re-
Beijing, China, and the Ph.D. degree in electronic search Assistant and Visiting Lecturer in the Depart-
engineering from Katholieke University Leuven, ment of Electrical Engineering and Computer Sci-
Belgium, in 1996, 1999, and 2006, respectively. ence, University of California, Berkeley. From 1991
From 1999 to 2000, he was an ASIC Engineer at to 1993, he was a Postdoctoral Research Assistant
Huawei High-Tech Incorporated, Beijing, and then a of the Belgian National Fund of Scientific Research,
Chip Design Engineer at Capella Microsystem Inco- ESAT Laboratory, Katholieke Universiteit Leuven. In
poration, Beijing. Currently, he is with Mindspeed Technologies, Inc, Newport 1993, he was appointed as a tenure Research Associate of the Belgian National
Beach, CA. His research interests are in the area of analog/mixed-signal inte- Fund of Scientific Research and also an Assistant Professor at the Katholieke
grated circuits design and computer-aided design tools. Universiteit Leuven. In 1995, he was promoted to Associate Professor at the
same university, where he is now a full-time Professor. His research interests
are in the design of analog and mixed-signal integrated circuits, and especially
in analog and mixed-signal computer-aided design tools and design automation
(modeling, simulation and symbolic analysis, analog synthesis, analog layout
generation, analog and mixed-signal testing). He is coordinator or partner of
several (industrial) research projects in this area. He has authored or coauthored
one book and more than 250 papers in edited books, international journals, and
conference proceedings.
Dr. Gielen served as the 2005 President of the IEEE Circuits and Systems
(CAS) Society. He was an Associate Editor of the IEEE TRANSACTIONS ON
CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS, and
later of IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: ANALOG AND
DIGITAL SIGNAL PROCESSING. He is also a member of the Program Commit-
tees of international conferences (ICCAD, DATE, ISCAS). He received the Best
Paper Award of the Wiley International Journal on Circuit Theory and Appli-
cations in 1995. He was the 1997 Laureate of the Belgian National Academy of
Sciences, Literature and Arts. He received the 2000 Alcatel Award of the Na-
tional Fund of Scientific Research for innovations in telecom research, and the
2004 Best Paper Award at the DATE conference.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 53, NO. 1, JANUARY 2006 3
Fig. 2. Voltage variation of the internal node. (a) Schematics. (b) Small-signal A. Second-Order Distortion Caused by the Delay Differences
equivalent circuit.
on the Clock Net
For a current-steering DAC without RZ output stage,2 the
impedance of the switch transistor. Thus, the variation of the output of the current sources are added directly to the output
output voltage will cause a small variation of , which will current of the DAC; the delay differences among these current
result in a change of the next switching time. Fig. 2(b) shows sources will reflect on the output current directly.
the small-signal equivalent circuit. The transistor which is off is Consider one of the current sources with clock delay .
omitted. When the output impedance of the (cascoded) current Assume it is switched on in the th sampling cycle. Then its
source is very large, with KCL the equation below can be output current is
obtained
(1)
(2)
where and are the transconductance and the output
impedance of the switch transistor respectively. This is the where is the sampling period, is the time constant decided
physical reason of the output-dependent delay differences. by the DACs load. The amplitude is set to be one. This is a
As the first part (part I) of our study, this paper will only simplified expression, the internal poles of the current cell have
focus on the analysis and improvement of the cell-dependent been omitted. But this is sufficient to get meaningful results.
delay differences. However, our calculations show that the same The distortion of this current source is thus
method can also be applied to the analysis of the output-depen-
dent delay differences to get meaningful results. The analysis
and improvement of the output-dependent delay differences will (3)
be proposed in the part II of this paper.
This paper is organized as follows. In Section II, the method Assume the DACs input is a sinusoidal signal
used in [9] is explained in full detail to analyze the impact of (4)
the delay differences in the clock net on the SFDR of the DAC.
Formulas with clear physical meaning are derived, and are com- where is the DACs number of bits. The amplitude is set so
pared to the measurement results from a published paper [1]. that the amplitude of every unit current source is 1.
Next, this method is extended in Section III to analyze the delay 2Current-steering DACs with RZ output stage will be analyzed later in Sec-
differences in the output net. The impact of the return-to-zero tion IV.
CHEN AND GIELEN: CURRENT-STEERING DACs DYNAMIC SFDR I 5
For a high-resolution DAC, ignoring the discrete nature of the cells are switched on in the same order as their delay increases
output signal, the ideal output of the DAC in the period from (referred to as linear switching sequence or LSS assumption).
to can be expressed as The LDDV assumption does not exist in a real DAC, but as will
be seen in Section V, it has little impact on the result, so it is
(5) justifiable to adopt this assumption to simplify the derivation.
(6) What will impact the SFDR is actually the distribution of
the delay values rather than the values themselves. The delay
During this period the DACs total distortion is distribution is determined by the switching sequence of the
switch-and-latch cells3 when the input code step by step changes
from the minimum value to the maximum value. Here, the LSS
assumption is one of the possible cases in a real DAC and it
leads to an interpretable analytic result. For the cell-dependent
(7) delays, when the positions of the switch-and-latch cells in the
switch-and-latch block are decided, the delay values for the
corresponding current cells are fixed, but the designer can still
where (3) has been used, and
decide the delay distribution freely by arranging the actual
when switching sequence of the switch-and-latch cells. We will see
(8)
else later from the simulation results in Section VI that the LSS
case is one of the worst delay distributions.4 However, through
is a square function.
the analysis of this worst case, we will obtain the optimized
In order to simplify the calculations, only linearly-distributed
delay distribution and the corresponding optimized switching
delays are considered for now. That is
sequence of the switch-and-latch cells.
(9)
3Note that the switching sequence of the switch-and-latch cells can be
where is a constant. Two assumptions are contained in this different from the switching sequence of the current source cells.
4Actually, what we simulated and optimized in Sections V and VI is
approximation. First, the delay values are integer times of .
the distribution of the delays in the output net. But we will see later that
(We will refer to this condition as linearly distributed delay the delay differences in both the clock net and the output net impact the
values or simply LDDV assumption later.) Second, the current SFDR in a similar way [see (23) and (46)].
6 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 53, NO. 1, JANUARY 2006
Using (9) into (7), with a little calculation we can get the result For the high-accuracy DACs of nowadays, it normally holds that
for the DACs total distortion in the th time window
(19)
(10) It meas that, the output settles well at the end of each sampling
cycle.
Define the function as Thus, (18) can be simplified into
(11) (20)
As will be seen later, determines the amplitudes of the With (16), (17), and (20), the Fourier transform of can
distortion components at different frequencies. Then (10) can be obtained
be simplified to
(12)
(23)
(24)
(17)
Then we can get by calculating and , respec-
tively.
Define the function as
(18)
5Results later on confirm that the second-order distortion is the dominating
(25)
contribution if the signal frequency is not too high.
CHEN AND GIELEN: CURRENT-STEERING DACs DYNAMIC SFDR I 7
where the square function is defined in (8). Then Observing (27) and (30), the relation holds
can be expressed as (ignoring the discrete nature of the output
signal) (34)
(26)
(36)
(27)
Thus, we have obtained the amplitude of the signal frequency.
With (16), and (27), s Fourier transform is obtained Combining (23) and (36), the SFDR of a current-steering
DAC without RZ output stage due to cell-dependent delay dif-
ferences caused by the clock net of the latches can be obtained
as (assuming that the second-order distortion is dominant)
(37)
(28)
Now we calculate the Fourier transform of . It can be where is the maximum delay difference of the
expressed as current cells. We see from this expression that the SFDR will
increase with decreasing delay differences ( ) of the current
cells. This is consistent with our intuition. The most interesting
thing about this result is the dependency of the SFDR on the
signal frequency and the sampling frequency. Fig. 4(a) shows
(29) the SFDR- curve, where the signal frequency
has been normalized to the sampling frequency .
Applying the Fourier transform When the signal frequency increases from zero up to , the
SFDR will first decrease and then increase. When ,
the SFDR reaches its lowest value. This result will be verified
with more detailed simulation results later in Section V.
(30) The SFDR- curve is shown in Fig. 4(b), where the sam-
With the above equations, together with (16), we get the Fourier pling frequency has been normalized to the signal frequency
transform of as . We conclude from this figure that, when the signal fre-
quency is a constant, we can improve the SFDR by decreasing
the sampling frequency. This is reasonable because, when the
sampling frequency increases, the distortion in every sampling
cycle appears with a higher frequency, thus deteriorating the
DACs SFDR property. When the sampling frequency becomes
(31)
even higher, the amplitude of the distortion in every sampling
cycle is reduced, the total distortion will not increase much,
From (24), (28), and (31), the Fourier transform of can
thus, the SFDR will approach a constant value, as shown in the
be obtained
figure.
(32) When the signal frequency is so low that , (37) can
be simplified into
The amplitude of the component at the signal frequency
is (38)
This result shows that the SFDR will decrease with increasing
signal frequency at a rate of about dB when the signal
(33) frequency is low. This is consistent with [1]’s measurement (see
8 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 53, NO. 1, JANUARY 2006
delay values extracted from the layout obtain a result very close
to this number. The reason will be analyzed in that section. Ob-
viously, the cell-dependent delay difference may indeed deteri-
orate the DACs SFDR property seriously.
(40)
where is the variation of the time constant , and is much
less than itself, say, . With this condition, (40) can be
simplified into
(41)
Since
we have
Fig. 4. Equation (37): the dependence of the SFDR on the signal frequency
and the sampling frequency due to the delay differences on the clock net for a
NRZ DAC. (a) SFDR-f curve. (b) SFDR-f curve.
So, (41) can be further simplified into
[1, Fig. 16]). Another conclusion which can be drawn from (38)
is that, when the signal frequency is low enough, the DACs
SFDR property will have nothing to do with the sampling fre- (42)
quency. In [1, Fig. 18], we can see that the DACs SFDR property
is nearly constant provided that the sampling frequency is lower As in Section II-A, the total distortion in the th sampling
than 150 MHz. (If the sampling frequency becomes higher, the cycle is
DAC will not be able to achieve 14 bits accuracy, thus the SFDR
will decrease.) This result also verifies our analytic result.
With (38) we can estimate the maximum signal frequency a
DAC can achieve under a given maximum delay difference. For
an N-bit DAC, the SFDR it should achieve is at least (43)
dB [10]. So the relation below should be satisfied
where and are defined in (5) and (6), and the function
is defined in (8).
Again, we assume the variation of the time constant, , to be
(39)
linearly distributed
For example, for a 14-bit DAC, when ps (extracted
from the layout of a real DAC [1]), the maximum frequency it (44)
can achieve is only 1.39 MHz, even when all other nonlinearities
are omitted. Of course, this result is based on the simplification where is a constant. The LDDV and LSS assumptions are in-
of (9). However, our simulations in Section V with the actual cluded in this equation just as in (9). Their impact on the results
CHEN AND GIELEN: CURRENT-STEERING DACs DYNAMIC SFDR I 9
Fig. 6. RZ DACs output signal. (a) The simplified output signal f (t). (b)
The actual output signal f (t).
Fig. 5. Equation (46): the dependence of the SFDR on the signal frequency
and the sampling frequency due to the delay differences on the output net for a
IV. IMPACT OF RZ OUTPUT STAGE ON DELAY DIFFERENCES
NRZ DAC. (a) SFDR-f curve. (b) SFDR-f curve. In [2] and [3], an RZ output stage was used to enhance the
SFDR property by setting the output to a fixed value (ac ground)
is also similar to the case of the delay differences in the clock at the start of any transition. With this architecture, the output
net, which has been discussed when we introduce (9). signal is divided into two phases in every clock cycle: the track
With similar calculations as in Section II-A, we can get the phase and the attenuate phase. The output tracks the DAC output
amplitude of the second-order distortion as only during the track phase, and it is attenuated to a very low
fixed value during the attenuate phase [2]. When the signal fre-
(45) quency increases, the SFDR property of such a DAC will de-
crease with a much slower rate compared to DACs without RZ
Combining (36) and (45), the SFDR of a DAC without output output stage. In this section, a simplified ideal RZ output stage
stage due to output time constant differences is obtained as with attenuation time assumed to approach zero as shown in
Fig. 6(a) will be analyzed as a worst-case situation to a real RZ
DAC. The impact of the cell-dependent delay differences on the
SFDR of such a DAC will be obtained. Since the delay differ-
(46) ences on the clock net and the output net impact the SFDR in a
where is the maximum time constant variation. This re- similar way, only the delay differences on the clock net are an-
sult is shown in Fig. 5. It is very similar to (37)’s result (see alyzed in this section.
Fig. 4), i.e., the impact of the output time constant variation on Fig. 6(a) shows the simplified output signal of an RZ DAC. In
the SFDR is similar to that of the clock delay difference. That this figure, the curve is the output of the RZ DAC. For com-
means that the output time constant variation will impact the parison, a normal (without RZ output stage) DACs output signal
SFDR in the same way as the clock delay difference does. Thus, is also shown. Both the signal amplitude at frequency
10 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS, VOL. 53, NO. 1, JANUARY 2006
(47)
With calculations similar to Section II-B, the amplitude of the
component at frequency can be obtained as
(48)
where
(49)
The second-order distortion of such a DAC has to be recalcu-
lated too. Instead of (7), the total distortion of the th cycle is
now
(50)
Fig. 11. Comparison of SFDR calculation and simulation results for a DAC
without output stage [1].
in real DACs. The reason why the simulation results fit the cal-
culation result is that, in the chip [1] we used, the current cells
are switched on nearly in the same order as their delays decrease Fig. 12. Reduction of the cell-dependent delay differences. (a) Delay
[Fig. 13(a)], i.e., it approximately fits the LSS assumption de- cancellation between the clock net and the output net. (b) Tree-like clock net
and output net.
fined in Section II-A. In such a case, the peak of the nonlinear
distortion caused by the different delays appears twice in one
sinusoidal cycle, and results in big second-order distortion, thus to what has been done to the current source cells in order to
deteriorating the SFDR. achieve good INL static property in [1] and [11]. Fig. 1 shows
We will show in Section VI by simulations that the delay dis- the floorplan of a typical high-accuracy current-steering DAC.
tribution (physically the switching sequence of the switch-and- The connections between the thermodecoder block and the
latch cells) greatly affects the DACs SFDR property assuming switch-and-latch block (connections 1 in the figure), together
that the delay difference values are fixed. Since it is hard for with the connections between the switch-and-latch block and
a high-accuracy DAC to reduce the delay differences of all the the current source block (connections 2), provide enough
current cells through proper design and layout, finding a best freedom to the designers for realizing in the same chip both
delay distribution becomes a promising way to solve the SFDR the optimum switching sequence in the current-source block
problem. for good static INL performance and the optimum switching
sequence in the switch-and-latch block for good dynamic
SFDR performance.
VI. DDC TECHNIQUE
If we only take the distortion caused by the delay differences
As aforementioned, we can reduce the delay differences by into consideration, two conclusions are justifiable based on the
making the delay differences on the clock net and those on the above analysis and simulations. First, to first-order approxima-
output net cancel with each other, as shown in Fig. 12(a). But tion, the total distortion is the sum of all the distortions in every
the length differences in the clock net and in the output net will sampling cycle; in each cycle, the distortion is the sum of the
not result in exactly the same delay differences, therefore this distortions of every current cell that is switched on; and the dis-
method cannot solve the problem completely. Another possible tortion of a current cell in a sampling cycle is linearly propor-
solution is to use a tree-like connection for both the clock net tional to the delay difference value of the current cell [see (3)
and the output net as shown in Fig. 12(b). This method will and (42)]. Second, the second-order distortion dominates when
inevitably increase the area and slow down the sampling fre- the signal frequency is not too high. Correspondingly, there are
quency, and for high-accuracy DACs, the tree-like buses will two solutions to improve the SFDR property: to reduce the am-
make it difficult to get a reasonable aspect ratio. plitude of the distortion in each sampling cycle, or to reduce the
The methods mentioned above work by reducing the values energy of the second-order distortion. Thus, we get two rules
of the delay differences. In this section, we will present another (we call them “the DDC rules”) for arranging the switching se-
method which will not change the delay difference values of the quence of the switch-and-latch cells
current cells. Instead, we will reduce the impact of the cell-de- 1) The amplitude of the distortion in every sampling cycle
pendent delay differences on the SFDR directly by properly should be reduced to as low as possible. That means that
choosing the switching sequence of the switch-and-latch cells. the current cells with big delay values should neighbor
In this way, what is changed is the delay distribution instead of the current cells with small delay values in the switching
the values of the delay differences. It is to some degree similar sequence.
CHEN AND GIELEN: CURRENT-STEERING DACs DYNAMIC SFDR I 13
TABLE I
POSITIONS OF SWITCH-AND-LATCH CELLS IN BLOCK
TABLE II
OPTIMIZED POSITIONS OF SWITCH-AND-LATCH CELLS IN BLOCK
Fig. 14. Result of optimizing the switching sequence for a real design case
(f = 150 MHz).
[2] A. R. Bugeja and B. Song, “A self-trimming 14-b 100-MS/s CMOS Georges G. E. Gielen (S’87–M’92–SM’99–F’02)
DAC,” IEEE J. Solid-State Circuits, vol. 35, pp. 1841–1852, 2000. received the M.Sc. and Ph.D. degrees in electrical en-
[3] J. Hyde, T. Humes, C. Diorio, M. Thomas, and M. Figueroa, “A 300- gineering from the Katholieke Universiteit Leuven,
ms/s 14-bit digital-to analog converter in logic CMOS,” IEEE J. Solid- Leuven, Belgium, in 1986 and 1990, respectively.
State Circuits, vol. 38, pp. 734–740, 2003. From 1986 to 1990, he was appointed as a Re-
[4] A. Van den Bosch, M. Borremans, M. Steyaert, and W. Sansen, “A 12 b search Assistant by the Belgian National Fund of
500 msample/s current-steering CMOS D/A converter,” in Proc. IEEE Scientific Research. In 1990, he was appointed a
2001 ISSCC, vol. XLIV, Feb. 2001, pp. 366–367. Postdoctoral Research Assistant and Visiting Lec-
[5] Y. Cong and R. Geiger, “A 1.5 v 14 b 100 MS/s self-calibrated DAC,” turer with the Department of Electrical Engineering
in Proc. IEEE 2003 ISSCC, vol. 1, Feb. 2003, pp. 128–482. and Computer Science, University of California,
[6] A. Van den Bosch, M. Steyaert, and W. Sansen, “SFDR-bandwidth lim- Berkeley. From 1991 to 1993, he was a Postdoctoral
itations for high speed high resolution current steering CMOS D/A con- Research Assistant with the Belgian National Fund of Scientific Research,
verters,” in Proc. IEEE 1999 ISCAS, 1999, pp. 1193–1196. ESAT Laboratory, Katholieke Universiteit Leuven. In 1993, he was appointed
[7] S. Luschas and H. S. Lee, “Output impedance requirements for DAC’s,” as a tenure Research Associate of the Belgian National Fund of Scientific
in Proc. IEEE 2003 ISCAS, May 2003, pp. I–861–I–864. Research and also an Assistant Professor at the Katholieke Universiteit Leuven.
[8] T. Chen and G. Gielen, “Modeling of the impact of the current source In 1995, he was promoted to Associate Professor at the same university, where
output impedance on the SFDR of current-steering CMOS DA con- he is now a full-time Professor. His research interests are in the design of analog
verters,” in Proc. IEEE 2004 ISCAS, May 2004. and mixed-signal integrated circuits, and especially in analog and mixed-signal
[9] , “Analysis of the dynamic SFDR property of high-accuracy cur- CAD tools and design automation (modeling, simulation and symbolic analysis,
rent-steering D/A converters,” in Proc. IEEE 2003 ISCAS, May 2003, analog synthesis, analog layout generation, analog and mixed-signal testing).
pp. I–973–I–976. He is coordinator or partner of several (industrial) research projects in this area.
[10] S. Haykin, An Introduction to Analog and Digital Communica- He has authored or coauthored one book and more than 250 papers in edited
tions. New York: Wiley, 1989. books, international journals, and conference proceedings.
[11] Y. Cong and R. Geiger, “Switching sequence optimization for gradient Dr. Gielen was the 2004 President-Elect of the IEEE Circuits and Systems
error compensation in thermometer-decoded DAC arrays,” IEEE Trans. (CAS) Society. He has been an Associate Editor of the IEEE TRANSACTIONS
Circuits Syst. II, Analog Digit. Signal Process., vol. 47, no. 7, pp. ON CIRCUITS AND SYSTEMS—I: FUNDAMENTAL THEORY AND APPLICATIONS,
585–595, Jul. 2000. and recently also of the IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS––II:
ANALOG AND DIGITAL SIGNAL PROCESSING. He is also a member of the Pro-
gram Committees of international conferences (ICCAD, DATE, ISCAS). He
received the Best Paper Award of the Wiley International Journal on Circuit
Tao Chen (S’02) was born in Zhangping City, Fujian Theory and Applications in 1995. He was the 1997 Laureate of the Belgian Na-
Province, P.R. China, in 1974. He received the B.S. tional Academy of Sciences, Literature, and Arts, and is a member of the Edi-
degree in electronic engineering in 1996 from Zhe- torial Board of the Kluwer International Journal on Analog Integrated Circuits
jiang University, and in 1999, he received the M.S. and Signal Processing.
degree, also in electronic engineering, from Tsinghua
University, China.
From 1999 to 2000, he was an ASIC engineer at
Huawei High-Tech Incoporation, Beijing, and then a
chip design engineer with Capella Microsystem In-
coporation, Beijing. Currently, he is a research as-
sistant with the MICAS Laboratories, K.U. Leuven,
Belgium. He is working toward the Ph.D. degree on the design of analog/mixed-
signal integrated circuits.