You are on page 1of 102

A Capacitance-Based Reference

Scheme for a 14b-Linear, 100


MS/s SAR-Assisted Pipeline ADC

Iniyavan Elumalai
Master of Science Thesis
A Capacitance-Based Reference
Scheme for a 14b-Linear, 100 MS/s
SAR-Assisted Pipeline ADC

Master of Science Thesis

For the degree of Master of Science in Microelectronics at Delft


University of Technology

Iniyavan Elumalai

August 21, 2012

Faculty of Electrical Engineering, Mathematics and Computer Science · Delft University of


Technology
Copyright © Broadcom Corporation
All rights reserved.

Copyright © Electrical Engineering


All rights reserved.
Delft University of Technology
Department of
Electrical Engineering

The undersigned hereby certify that they have read and recommend to the Faculty of
Electrical Engineering, Mathematics and Computer Science for acceptance a thesis
entitled
A Capacitance-Based Reference Scheme for a 14b-Linear, 100 MS/s
SAR-Assisted Pipeline ADC
by
Iniyavan Elumalai
in partial fulfillment of the requirements for the degree of
Master of Science Microelectronics

Dated: August 21, 2012

Supervisor(s):
Prof. Dr. John R. Long

Dr. Klaas Bult

Christopher Ward

Reader(s):
Prof. Dr. Kofi Makinwa

Dr. Wouter A. Serdijn

Dr. Frank Van der Goes


Abstract

Voltage reference buffers have always been the most power-hungry blocks in switched-capacitor
SAR ADCs. High frequency dynamic loading of the buffer by the capacitive DAC causes
glitches on the reference voltage, and the buffer has to be fast enough to recover such tran-
sients to 1/2 LSB precision in every bit cycle. Such stringent requirements entail an immensely
capable buffer for high-speed, high-resolution converters. One obvious solution circumventing
the active buffer is to use a stand-alone capacitor that holds the reference voltage during SAR
evaluation. This solution accompanies, besides good power efficiency, excellent DAC settling
speed and noise properties, thereby greatly helping the ADC FoM. Reducing the capacitance
of the stand-alone capacitor for chip area concerns brings in code-dependent errors on the
reference, leading to heavy distortion.
This project implements an area-conscious, capacitance-based reference buffer scheme for
a 14b-linear, 100 MS/s SAR-assisted pipeline ADC in 28 nm CMOS. With moderately-valued
capacitances, two elaborate calibration techniques are proposed that help to suppress the
reference-induced distortion to less than 84 dB, effectively not degrading the SNDR. Designed
for 12-bit SNR, the prototype ADC with the implemented reference scheme consumes
1.81 mW and achieves a Schreier FoM of 175.5 dB.

Master of Science Thesis i Iniyavan Elumalai


ii Abstract

Iniyavan Elumalai Master of Science Thesis


Acknowledgments

It feels good to finish a thesis. It feels better to reminisce the times, and thank everyone who
had been of help.
I would like to begin by thanking Dr. Klaas Bult, who, besides introducing me into
the world of circuits with colourful, entertaining slides, provided me with this wonderful
opportunity to work at Broadcom. His support throughout this project is invaluable, and his
knowledge and experience in circuit design will continue to motivate my aspirations.
My supervisors at Broadcom, Christopher Ward and Dr. Frank Van der Goes, had been
of excellent guidance during this project. Chris, with his knack for animated explanations
that clarify even the most obscure of concepts, and Frank, with his never-ending stream of
innovative ideas will remain great sources of inspiration to me.
I would also like to thank my promoter, Prof. John Long for his extensive support. If
the reader finds this thesis uncluttered, if this writing has been at all effective, it is because
of his elaborate and meticulous feedback.
I want to thank Prof. Kofi Makniwa and Dr. Wouter A. Serdijn for teaching me more
of circuits, and am glad that they are serving on my thesis committee.
I am grateful to all the engineers at Broadcom Bunnik for the good times in the office,
and Els Van Zijl for taking good care of me and most of my administrative tasks. I would
like to acknowledge Rohan Sehgal and Nitz Saputra for providing guidance in improving my
thesis.
I owe the best times of my graduate life to my friends, in particular, Vamshi, Sangan,
Mandi, Anna and Jort, who have respectively got me through an arduous year of mount-
ing assignments, minor financial troubles, winter, summer and a precocious post-graduate
delirium.
To my family - amma, appa and akka, thank you for your love and other happy things
in my life.

Master of Science Thesis iii Iniyavan Elumalai


Table of Contents v

Table of Contents

Abstract i

Acknowledgments iii

1 Introduction 1
1-1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1-2 Limitations of SAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1-2-1 SAR Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1-2-2 Requirements on the Reference Buffer . . . . . . . . . . . . . . . . . . 6
1-3 Research Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1-4 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Speed Limitations in SAR-based A/D Converters 9


2-1 Basic Nonidealities in Charge Redistribution SAR A/D Converters . . . . . . 9
2-1-1 Quantization Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2-1-2 Thermal Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2-1-3 Static Nonidealities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2-1-4 Dynamic Nonidealities . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2-2 Dynamic Performance Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2-3 Speed Limitations of a SAR ADC . . . . . . . . . . . . . . . . . . . . . . . . . 15
2-4 Popular Techniques in Relevant Literature . . . . . . . . . . . . . . . . . . . . 15
2-4-1 Asynchronous Clocking . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2-4-2 Overranging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2-4-3 Pipelining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2-5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Master of Science Thesis Iniyavan Elumalai


vi Table of Contents

3 Proposed Reference Scheme 21


3-1 Conventional Reference Circuitry . . . . . . . . . . . . . . . . . . . . . . . . . 21
3-2 A Stand-alone Capacitor as the Reference Buffer . . . . . . . . . . . . . . . . 25
3-3 Deterministic Errors on the Reference Voltage . . . . . . . . . . . . . . . . . . 26
3-4 Benefits of the Proposed Reference Scheme . . . . . . . . . . . . . . . . . . . 31
3-5 Design Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3-6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Charge-based Compensation of Reference Errors 35


4-1 Reference-induced Distortion . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4-2 Error Compensation in Charge Domain . . . . . . . . . . . . . . . . . . . . . 38
4-3 Implementation Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4-3-1 Systematically Decreasing Vref . . . . . . . . . . . . . . . . . . . . . . 42
4-3-2 Cancelling Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4-3-3 Compensating for a Fewer Cycles . . . . . . . . . . . . . . . . . . . . . 45
4-4 Sources of Inaccuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4-4-1 Capacitor Mismatch . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4-4-2 Switch Nonidealities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4-5 Reference Scheme Implementation in the Fine SAR ADC . . . . . . . . . . . 47
4-6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4-7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Digital Correction of Reference Errors 51


5-1 Need for Digital Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5-2 Calibration Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5-2-1 Residue Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5-2-2 Calibration Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5-2-3 Deterministic Calibration Procedure . . . . . . . . . . . . . . . . . . . 57
5-3 Reference Scheme Implementation in Coarse SAR ADC . . . . . . . . . . . . 58
5-3-1 System-level Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5-3-2 Circuit-level Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5-4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5-5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Conclusion 71
6-1 Thesis Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6-2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

A Code-dependent Charge Steering from the Reference Source 75


A-1 Set-and-down Switching Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 76
A-2 Split-capacitor Switching Algorithm . . . . . . . . . . . . . . . . . . . . . . . 77

Bibliography 79

Iniyavan Elumalai Master of Science Thesis


List of Figures vii

List of Figures

1-1 State-of-the-art ADCs with their energy efficiency . . . . . . . . . . . . . . . . . 2


1-2 State-of-the-art ADCs with their Schreier FoM . . . . . . . . . . . . . . . . . . . 3
1-3 Simplified schematic of a charge redistribution SAR ADC . . . . . . . . . . . . . 4
1-4 Capacitive DAC configurations during different phases of operation for the 4-bit
digital code: 0110 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1-5 DAC voltage during the conversion phase for the 4-bit digital code: 0110 . . . . 6

2-1 Quantization error in an ideal A/D converter . . . . . . . . . . . . . . . . . . . . 10


2-2 Thermal noise in a charge redistribution SAR ADC . . . . . . . . . . . . . . . . 11
2-3 Differential and integral non-linearity in an ADC . . . . . . . . . . . . . . . . . . 12
2-4 Effect of incomplete DAC settling on the conversion result . . . . . . . . . . . . 13
2-5 Dynamic performance of a non-ideal 14-bit ADC . . . . . . . . . . . . . . . . . 14
2-6 Timing in a synchronous clocking scheme . . . . . . . . . . . . . . . . . . . . . 15
2-7 Timing in an asynchronous clocking scheme . . . . . . . . . . . . . . . . . . . . 16
2-8 SAR conversion illustrating easy and hard decisions . . . . . . . . . . . . . . . . 17
2-9 Effect of incomplete DAC settling on the conversion result in the presence of
overranging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2-10 A two-stage, SAR-assisted pipeline ADC . . . . . . . . . . . . . . . . . . . . . . 19

3-1 Code-dependent loading of reference source . . . . . . . . . . . . . . . . . . . . 21


3-2 Code-dependent transients on Vref recovered by the charge supplied by the buffer 22
3-3 Surface plot of the current required in the reference buffer versus sampling fre-
C
quency and the ratio Cref
dac
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3-4 (a) Reference capacitor disconnected during SAR conversion (b) Code-dependent,
deterministic errors on Vref . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3-5 3-bit set-and-down switching procedure . . . . . . . . . . . . . . . . . . . . . . 27

Master of Science Thesis Iniyavan Elumalai


viii List of Figures

3-6 Switching scenario in bit cycle i = 1 . . . . . . . . . . . . . . . . . . . . . . . . 28


3-7 Two different switching scenarios in bit cycle i = 0 . . . . . . . . . . . . . . . . 28
3-8 Different possible reference charges for a 3-bit set-and-down switching procedure 29
3-9 Possible values of reference charge per step for an 8-bit set-and-down switching
procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3-10 Total reference charge versus output code for an 8-bit set-and-down switching
procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3-11 Two-stage, SAR-assisted pipeline ADC . . . . . . . . . . . . . . . . . . . . . . . 32

4-1 Code-dependent errors on Vref = 1 V in the case of 8-bit set-and-down switching


with Cref = Cdac (simulated in MATLAB) . . . . . . . . . . . . . . . . . . . . . 36
4-2 Reference voltage at the end of the Convert phase versus 8-bit output code (sim-
ulated in MATLAB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4-3 Performance of an actual 8-bit ADC with binary-weighted DAC and stand-alone
reference capacitor, Cref = Cdac . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4-4 Performance of an actual 8-bit ADC with binary-weighted DAC and stand-alone
reference capacitor, Cref = 10Cdac . . . . . . . . . . . . . . . . . . . . . . . . . 37
4-5 Conceptual illustration of error compensation in charge domain . . . . . . . . . . 38
4-6 Code-dependent errors on Vref in the case of 5-bit set-and-down switching . . . 39
4-7 Illustration of compensation in bit cycle i = 3 . . . . . . . . . . . . . . . . . . . 40
4-8 Illustration of compensation in bit cycle i = 2 . . . . . . . . . . . . . . . . . . . 41
4-9 Conceptual illustration of reference compensation loop (in grey) along with the
main SAR loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4-10 Compensation for (a) constant Vref , and (b) code-independent, systematically
decreasing Vref . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4-11 Effect of Vref dropping to ′
Vref on effective signal coverage (a) with conventional
capacitive division ratio, and (b) with appropriately changed capacitive division . 43
4-12 Removing the code-dependent spread in Vref (a) by adding charge, and (b) by
removing charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4-13 Effect of DAC capacitor mismatch on the code-dependent errors in Vref (simulated
in MATLAB) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4-14 Single-ended schematic of the capacitive DAC in the fine stage . . . . . . . . . . 48
4-15 Examples of positive and negative DAC voltage waveforms in the fine stage . . . 48
4-16 Effective signal coverage per bit cycle with the designed cap. radix, taking into
account the code-independent drops in Vref and nominal switch charge injection
in every bit cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4-17 Fine ADC performance with the designed reference scheme including charge-based
compensation in 3 bit cycles, and changed capacitive radix to accommodate the
code-independent drop in Vref . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5-1 Ideal DAC tree for a 2-bit, differential search . . . . . . . . . . . . . . . . . . . . 53


5-2 Ideal residue characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5-3 Effect of reference errors on the DAC tree . . . . . . . . . . . . . . . . . . . . . 54

Iniyavan Elumalai Master of Science Thesis


List of Figures ix

5-4 Effect of reference errors on the residue characteristic . . . . . . . . . . . . . . . 54


5-5 DAC voltage modulation in an additional step after SAR conversion . . . . . . . 55
5-6 DAC tree with three different cases of DAC voltage modulation . . . . . . . . . . 56
5-7 Determining h1 (j) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5-8 Determining h2 (j) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5-9 Timing for the two ADC lanes in the prototype during calibration . . . . . . . . 59
5-10 Simplified block diagram of the coarse ADC . . . . . . . . . . . . . . . . . . . . 59
5-11 Schematic of the latch in the main SAR loop . . . . . . . . . . . . . . . . . . . 60
5-12 Flipped voltage follower and its output impedance versus frequency . . . . . . . 61
5-13 Schematic of the implemented pre-charge buffer . . . . . . . . . . . . . . . . . . 63
5-14 Transient response of the implemented pre-charge buffer . . . . . . . . . . . . . 64
5-15 Allowed locations for the test voltages required during calibration . . . . . . . . . 64
5-16 Schematic of a unit current cell . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5-17 Schematic of the output network in the calibration DAC . . . . . . . . . . . . . 65
5-18 Transfer function of the implemented calibration DAC . . . . . . . . . . . . . . . 66
5-19 Worst-case INL/ DNL performance of the implemented calibration DAC . . . . . 66
5-20 INL of the capacitive DAC with Cref = 20Cdac before calibration . . . . . . . . . 67
5-21 Coarse ADC performance with Cref = 20Cdac before calibration . . . . . . . . . 68
5-22 INL of the capacitive DAC with Cref = 20Cdac after calibration . . . . . . . . . 68
5-23 Coarse ADC performance with Cref = 20Cdac after calibration . . . . . . . . . . 69

6-1 Schreier FoM plot of the state-of-the-art ADCs along with the prototype . . . . . 72

A-1 Positive and negative DAC steps . . . . . . . . . . . . . . . . . . . . . . . . . . 76


A-2 Illustration of set-and-down switching algorithm in step i . . . . . . . . . . . . . 77
A-3 Illustration of split-cap switching algorithm in step i . . . . . . . . . . . . . . . . 78

Master of Science Thesis Iniyavan Elumalai


x List of Figures

Iniyavan Elumalai Master of Science Thesis


List of Tables xi

List of Tables

5-1 SR inputs during different modes of operation . . . . . . . . . . . . . . . . . . . 60


5-2 Specifications for the calibration DAC . . . . . . . . . . . . . . . . . . . . . . . 65

Master of Science Thesis Iniyavan Elumalai


xii List of Tables

Iniyavan Elumalai Master of Science Thesis


Chapter 1

Introduction

At the time of writing this dissertation, it is evident that mobile handset usage has long ago
surpassed critical mass. Room for further sustenance lies in the efforts to deliver a service to
end users that ideally expect instant, on-the-go access to a multitude of wireless services from
their battery-powered handsets. In that regard, the recent past has witnessed standardization
efforts for future wireless usage that accommodate seamless data delivery at faster rates with
better quality of service. What with the ever-growing list of wireless usage scenarios ranging
from telephony and mobile internet to short-range communications and data tethering, the
device manufacturers have a hard time building and integrating interfaces to several such
standards in a single chipset. Achieving a unified, portable hardware platform that supports
simultaneous reception of multiple modes and services from a battery power source requires
rigorous research and innovation efforts from the industry.
Present-day 3G/4G receivers offer simultaneous reception by employing parallel radio
front ends, with the IF and baseband sections for different standards occasionally sharing
resources. The analog-intensive RF portions remain dedicated to each and every standard.
Despite offering flexibility to orthogonally optimize each radio and achieve low power, this
approach does not seem attractive in terms of size and cost for future multi-standard handsets.
An alternative, more promising realization would be the software-defined radio (SDR) as
envisaged by Mitola [1], where the signal is digitized at RF with an analog-to-digital converter
(ADC) close to the antenna, and the entire signal processing definitions are realized in the
completely reconfigurable, low-cost digital domain. This idealistic approach, however, remains
elusive as the specifications imposed on the ADC are not within sight at present.
For a radio platform covering all the current wireless standards from 800 MHz to 6 GHz
[2], Nyquist-sampled data conversion is required at a sampling speed above 12 GS/s. Fur-
thermore, since this approach does not proffer the radio with any tolerance to the interferers,
which could be as high as 100 dB larger than the desired signal, the ADC should also have
an extraordinary dynamic range. Extrapolating from a survey of currently existent data con-
verters [3] as shown by the dotted lines in Figure 1-1, such an ADC would dissipate around
1 kW - far from the tolerable level of 20 mW given currently available batteries - which is
clearly impractical in a portable handset. The requirements could only grow more stringent as

Master of Science Thesis 1 Iniyavan Elumalai


2 Introduction

time progresses, considering the inevitable increase in the number and strength of interferers.
Therefore, it becomes necessary to look at the current directions in ADC design, and work
out the significant challenges towards realizing a low-power, raw ADC located close to the
antenna, which will eventually enable frequency-agile cognitive radios for mobile handsets.
This research work is one such effort.

1-1 Motivation

Figure 1-1 presents the state-of-the-art ADCs published at the recent ISSCC and VLSI confer-
ences with their power efficiency, expressed as energy per Nyquist sample (P ower/fs , where
fs is the sampling clock frequency), logarithmically plotted against the ADC’s signal-to-noise
and distortion ratio (SNDR) in dB. The plot also includes the benchmark lines corresponding
to figures of merit (FoM) commonly used for gauging a converter’s energy efficiency.

1.E+06
ISSCC 2008-2013
1.E+05 VLSI 2008-2012

~ 1 kW for fs = 12 GHz
1.E+04

1.E+03 FOMS = 170 dB


P/fs [pJ]

1.E+02

1.E+01
FOMW = 10 fJ/conversion-step

1.E+00

1.E-01
10 20 30 40 50 60 70 80 90 100 110
SNDR at fnyq [dB]

Figure 1-1: State-of-the-art ADCs with their energy efficiency

The Walden FoM [4], given by

P ower
F oMW = , (1-1)
fs ∗ 2EN OB

normalizes the energy per Nyquist sample by the effective number of bits resolved (ENOB)
under the assumption that doubling the precision doubles the power consumption. While this
holds for low-to-medium resolution converters, as illustrated by a benchmark trend line of

Iniyavan Elumalai Master of Science Thesis


1-1 Motivation 3

10 fJ/conversion-step, high resolution converters (> 12 b) having to expend additional power


to alleviate thermal noise call for another figure of merit.
The Schreier FoM [5], given by
fs
F oMS = SN DR + 10 ∗ log , (1-2)
P ower
predicts a 4x increase per every added bit, after approximately 75 dB SNDR, where the
thermal noise limitation mostly supersedes that of quantization. The plot in Figure 1-2
displays Schreier FoMs of aforementioned ADCs against their Nyquist sampling rate, with
the dashed lines representing the limitation of presently available technology. Achieving the
ultimate ADC FoM for simultaneous full-band capture in a mobile radio front end, indicated
by the red asterisk in the top right of the plot, requires several careful steps, and might take
a few years benefiting from technology scaling.

190
Target
180

170

160
FOMS [dB]

150

140
SAR-based
130 Others

120

110
1.E+04 1.E+05 1.E+06 1.E+07 1.E+08 1.E+09 1.E+10 1.E+11
f s [Hz]

Figure 1-2: State-of-the-art ADCs with their Schreier FoM

Besides device scaling, moving the energy efficiency to higher speeds can be expected
from the popular technique of time-interleaving several ADC lanes that operate efficiently at
moderate speeds. The overhead to resolve the problems arising from interleaving, however,
counteracts the improvement in FoM. This necessitates research efforts in two directions:
(1) resolving the errors due to time-interleaving [6] with minimal power overhead, and (2)
improving the speeds of individual, energy-efficient ADC lanes. The latter approach fuels the
motivation of this research.

Master of Science Thesis Iniyavan Elumalai


4 Introduction

The empirical data show that converters based on the successive approximation prin-
ciple (SAR) promise excellent energy efficiency at moderate speeds (as evident in Figure
1-2). Boasting minimal analog complexity, the SAR architecture gets clear benefits from the
continuous scaling of digital circuitry, and seems well-suited to deep-submicron technologies.
However, since the successive approximation is performed by iterations of decision making,
the architecture falters in speed at high resolutions. The goal of this research is to relieve the
SAR architecture of its speed limitations without compromising on its power efficiency.

1-2 Limitations of SAR


Before discussing the various speed limitations in the SAR architecture, a brief overview of
its operation is given.

1-2-1 SAR Overview

The most popular implementation of the SAR A/D converter is a charge-redistribution SAR
[7], a simplified N-bit schematic of which is shown in Figure 1-3. The core of the converter,
comprising reference voltage circuitry, a binary comparator, SAR control logic, switches and
an array of radix-weighted capacitors (CN , CN −1 , ..., C0 ), implements a radix-based search
for the sampled input through all of the reference quantization levels before converging on
the digital output. The digital-to-analog conversion (DAC) is by virtue of the capacitor array
and switches, which successively approximates the input voltage. The conversion process in
a binary SAR is detailed next with an illustration of a 4-bit binary conversion.

Vdac − Vi
CN CN −1 C1 C0
Vref
Vi Switch array SAR Logic

Dout
Figure 1-3: Simplified schematic of a charge redistribution SAR ADC

The input signal Vi is first sampled onto the array of binary-weighted capacitors as shown
in Figure 1-4.a), after which the conversion process starts. The binary search begins at mid-
scale by comparing the acquired input voltage with one-half of the reference voltage. Halving
the reference at the comparator input is performed by connecting the MSB capacitor, whose
capacitance is equal to the sum of the remaining capacitances in the array, to the reference
voltage, and the rest to ground, as shown in Figure 1-4.b). The output of this comparison
represents the MSB decision.
The SAR control logic learns this decision, and knows which half of the signal range is
to be considered for the next comparison. For a positive decision, the SAR creates a decision

Iniyavan Elumalai Master of Science Thesis


1-2 Limitations of SAR 5

8 4 2 1 1

Vi
(a) Sampling

Vref
8
8Vref
16 − Vi
D3 = 0
4 2 1 1

(b) i = 3

Vref
4
4Vref
16 − Vi
D2 = 1
8 2 1 1

(c) i = 2

Vref
4 2
6Vref
16 − Vi
D1 = 1
8 1 1

(d) i = 1

Vref
4 2 1
7Vref
16 − Vi
D0 = 0
8 1

(e) i = 0
Figure 1-4: Capacitive DAC configurations during different phases of operation for the 4-bit
digital code: 0110

threshold at mid-scale in the upper half by connecting the MSB-1 capacitor to the reference
which drives Vdac from 1/2 Vref to 3/4 Vref. For a negative decision, the voltage has to be
driven in the negative direction by the same amount. This is performed as shown in Figure

Master of Science Thesis Iniyavan Elumalai


6 Introduction

1-4.c), by toggling the MSB capacitor from the reference to ground, and connecting the
MSB-1 capacitor to the reference. This comparison localizes the input to a quarter of the
signal range.
As the conversion process progresses with every decision being made, the effective signal
range (represented by a gray bar in Figure 1-5 for every bit cycle) reduces by a factor of two,
and the decision threshold for next comparison is derived from the reference by toggling the
top plate switches accordingly. This process proceeds in a loop: the comparator makes a
decision, the control logic determines the next decision threshold, the DAC settles to the cor-
responding voltage, and so on, until the algorithm converges at the last cycle. The difference
between Vdac and Vi is then less than one LSB, which represents the residual quantization
error. Figure 1-5 shows how the DAC voltage tracks the input during SAR conversion.

Vref

Vdac
Vref /2
<LSB

0 0 1 1 0
(D3 ) (D2 ) (D1 ) (D0 )
time
Figure 1-5: DAC voltage during the conversion phase for the 4-bit digital code: 0110

1-2-2 Requirements on the Reference Buffer

Refreshing the voltage (Vdac ) every bit cycle to derive different quantization levels from a single
reference requires that the reference source must supply charge throughout the process. The
charge required depends on from where, and to where the node is driven, which indicates the
strong dependence on the switching decisions being made throughout the whole conversion.
Pulling such signal dependent charge from the reference causes fluctuations in the buffered
reference voltage. However, the accuracy requirements on the reference voltage are stringent
at up to 1/2 LSB precision every bit-cycle, the settling time of which in turn fundamentally
limits the ADC conversion speed. It is therefore important to develop techniques that can
deal with such high-frequency disturbances.

Iniyavan Elumalai Master of Science Thesis


1-3 Research Contribution 7

Traditionally, power-hungry buffers and/or large decoupling capacitors are employed to


create a constant reference voltage from an off-chip bandgap. Using a fast active buffer implies
that the disturbances settle to the required precision within a bit cycle, that is, before the
next comparison is performed. In high-speed, high-resolution systems, this demands large
currents to be delivered with fast transitions. Voltage reference buffers (mostly employed
off-chip) have always been the most power-hungry blocks in switched-capacitor SAR ADCs
[8, 9]. Furthermore, as predicted in [10], device scaling exacerbates the power dissipation of
such noise-limited, settling-dominated circuits. Also, using large decoupling capacitors such
that the disturbances are reduced to within an LSB also does not seem to be viable in future,
because the bypass capacitance required is around 2N times the total array capacitance, where
N is the required system resolution.

1-3 Research Contribution

It can be clearly identified that the reference buffer places a serious limitation on the speed
of a SAR, which can only be overcome at the expense of power and area. A novel reference
buffer scheme suitable for high-end SAR converters is proposed in this research project that
improves the SAR speed without compromising power and area efficiency.

A stand-alone capacitor holding the reference voltage during conversion is an attractive


solution for fast settling. Reducing the capacitance for area concerns causes signal-dependent
errors on the reference voltage leading to heavy distortion. The basic idea underlying this
proposal is that those signal-dependent errors, as mentioned earlier, are completely determin-
istic given the knowledge of switching decisions being made and thus can be corrected. This
research explores two different techniques of dealing with reference-induced distortion.

With knowledge of the SAR decisions being made, one solution is to compensate the
charge on the reference capacitance by the same amount of charge being pulled from it every
bit cycle. This solution employs a bank of capacitors, from which an appropriate amount of
capacitance holding the required charge is connected to the main reference capacitor when
a switching decision is made that will compensate it accordingly. The other approach is
to allow errors in the analog domain, and then correct them downstream through a digital
calibration scheme. During calibration, the reference error for every digital code is estimated
by processing the corresponding residue voltage, which contains the required information
about the reference voltage at the end of that SAR conversion.

A two-stage, SAR-assisted pipeline ADC, designed by the research team in Broadcom


Netherlands, serves as the platform for this study. With two stages employing each of the
aforementioned reference buffer schemes, implementation choices for both of the calibration
techniques are studied in this dissertation. Reasonably-sized reference capacitances that are
not very much larger than the array capacitance will be used. This work attempts to maintain
the reference-induced distortion below 14-bit linearity with the aid of calibration techniques,
and demonstrates by simulations the improvement in speed and FoM therof.

Master of Science Thesis Iniyavan Elumalai


8 Introduction

1-4 Thesis Organization

The organization of this dissertation is as follows: Chapter 2 gives a detailed account of the
factors limiting the speed of a SAR ADC, and explains how relieving the settling-imposed
limitations translates to stringent requirements on the reference buffer. Chapter 3 states the
main proposal of this work: to replace the active reference buffer with a storage capacitor, and
briefly discusses the ensuing benefits and design problems associated with that. In Chapters
4 and 5, two different calibration schemes for correcting non-linearities due to errors on
reference voltage are proposed and elaborated. The circuit-level implementation details of
the proposed schemes on a two-stage pipelined SAR ADC are also presented along with the
simulation results. Chapter 6 concludes this dissertation and talks about the future scope for
work on this topic.

Iniyavan Elumalai Master of Science Thesis


Chapter 2

Speed Limitations in SAR-based


A/D Converters

In this chapter, we investigate some basic non idealities associated with charge redistribution
SAR A/D converters and review a number of performance measures commonly used for data
converters. Following this, factors limiting the SAR speed are identified and given a brief
treatment. Finally, we look at a few popular techniques in the recent literature that attempt
to relieve the SAR converter of its speed limitations.

2-1 Basic Nonidealities in Charge Redistribution SAR A/D


Converters

The quantization error, inherent in any data converter, represents the lowest limit on how
small the error power can be. Due to various circuit imperfections, the total error power is
always larger in an actual implementation. This section identifies the basic nonidealities in a
charge redistribution SAR A/D converter and attributes them to the error sources, which are
essentially those of the capacitive DAC, the comparator and the reference voltage circuitry.

2-1-1 Quantization Noise

The maximum error an ideal converter makes in digitizing an input signal is ± 1/2 LSB,
as shown in Figure 2-1(a). This quantization error, also referred to as the residue of the
conversion, plotted against input in Figure 2-1(b) is divided into 2N segments, and it changes
linearly with the input within a segment. For a high-resolution converter, it is reasonable
to assume that this error is uniformly distributed with equal probability at any value in the
range ± 1/2 LSB. The root-mean-square value of such a uniformly distributed error can be

derived to be equal to ∆/ 12 [11].

Master of Science Thesis 9 Iniyavan Elumalai


10 Speed Limitations in SAR-based A/D Converters

When viewed in the frequency domain, this sawtooth waveform produces harmonics
across the frequency band, which when folded back into the Nyquist bandwidth, sum together
to produce a quantization noise power of ∆2/12. The spectrum in the Nyquist bandwidth
is approximately Gaussian, except (under certain conditions,) when the quantization noise
becomes correlated with the input signal and power is then concentrated close to the signal
harmonics.

Digital
Output

Analog input

(a) Ideal A/D transfer function

+∆/2

Analog input
−∆/2

(b) Quantization error

Figure 2-1: Quantization error in an ideal A/D converter

2-1-2 Thermal Noise

Besides the inherent quantization noise, a charge-redistribution SAR ADC suffers from ther-
mal noise. All the intrinsic thermal noise sources, when referred to the input of the comparator
input, result in an equivalent noise voltage given by
kT
Vn,eq 2 = + Vn,comp 2 + Vn,ref 2 . (2-1)
Cdac
The thermal noise associated with the on-resistance of the switches, when sampled on Cdac ,
contributes the first term in (2-1), which is the total integrated noise in the Nyquist band-
width. The second term (Vn,comp 2 ) represents the input-referred thermal noise of the compara-
tor, and the last term (Vn,ref 2 ) is the noise contribution from the reference voltage circuitry.
Figure 2-2(a) shows these thermal noise contributions in a simplified representation.

Iniyavan Elumalai Master of Science Thesis


2-1 Basic Nonidealities in Charge Redistribution SAR A/D Converters 11

It should be noted that any noise on the reference voltage is multiplicative with the input
and adds to the total ADC noise floor as shown in Figure 2-2(b). This can be understood given
the basic principle upon which a SAR ADC works: performing approximation by multiplying
the input voltage with 2N /Vref . Therefore, any error on the reference voltage has a signal-
dependent gain towards the output. On a circuit level, it can be understood that the reference
noise is low-pass filtered by the DAC, where the bandwidth is signal-dependent. Thus, from
(2-1), the array capacitance should be adequately high in order to reduce the total equivalent
noise voltage to less than an LSB. Also, the comparator and reference circuitry must be
designed for appropriately low noise.

Vi

Rn,ref c ∗ Cdac Rn,comp


Vref

(1 − c) ∗ Cdac

Code-dependent
network

(a) Thermal noise contributions

Noise Reference noise

Comparator +
Sampling noise

Code

(b) Code-multiplicative reference noise

Figure 2-2: Thermal noise in a charge redistribution SAR ADC

2-1-3 Static Nonidealities

The common sources of static errors in a charge redistribution SAR ADC are component
mismatches in the capacitive DAC and the comparator. Degrading an ideal ADC transfer
function are offset, gain error and non-ideal code transition points.

The differential non-linearity (DNL) and integral non-linearity (INL) are often used as
performance measures for static linearity. DNL represents the deviation of the step size in a
non-ideal ADC from the ideal size ∆. If Xk is defined as the transition point between two

Master of Science Thesis Iniyavan Elumalai


12 Speed Limitations in SAR-based A/D Converters

consecutive codes k-1 and k, then the normalized DNL is given by


Xk − Xk−1 − ∆
DN L(k) = . (2-2)

The total deviation of the actual transfer function from a best-fit line (thus ignoring the
effects of offset and gain-error) is defined as the integral non-linearity. INL is expressed as
k
X
IN L(k) = DN L(l). (2-3)
l=1

Usually, a single number is mentioned for both INL and DNL for an ADC, which rep-
resents the worst deviation among all of the bins. While offset and gain-error are usually
acceptable and, if not, may be corrected through simple digital calibration schemes, the
linearity of the ADC transfer is indispensable. Since the matching properties of the array
capacitors crucially affect the linearity characteristics of an ADC, care has to be taken in the
design and layout of the capacitive DAC.

Digital
output

+0.5 LSB INL


Xk−1 Xk
Analog input


+0.5 LSB DNL

Figure 2-3: Differential and integral non-linearity in an ADC

2-1-4 Dynamic Nonidealities

Besides static errors due to capacitor mismatches, the linearity performance in a charge
redistribution ADC is affected by several other dynamic error sources that challenge the
accuracy of the inherent digital-to-analog conversion.

DAC Settling

Classified as an error source, this is rather a design requirement. As explained earlier in


Section 1-2-2, different reference levels are derived during conversion by connecting different
amount of capacitance to Vref or the ground. Doing so causes disturbances on the buffered

Iniyavan Elumalai Master of Science Thesis


2-2 Dynamic Performance Measures 13

reference voltage. The buffer recovers to establish the required DAC voltage before com-
parison, limited by its slew rate and settling capabilities. Figure 2-4 shows what happens if
the DAC settles incompletely. The MSB decision goes wrong, and thanks to the sequential,
irrevocable nature of binary search, the error cannot be compensated, thereby leading to an
incorrect digital output. Therefore, the DAC settling before every comparison should be ac-
curate to within 1/2 LSB error, for every decision should be as accurate as the final result of
the conversion [12].

Vref
15
14
13
12
11
10
9
8
Vdac Settling 7
error 6
5
4
3
2
1
0
0
1 0 0 0
(D3 ) (D2 ) (D1 ) (D0 )

time
Ideal decision threshold

Figure 2-4: Effect of incomplete DAC settling on the conversion result

Switch Nonidealities

As in any switched-capacitor circuit, nonidealities from MOS switches, namely, clock feedthrough,
channel charge injection and non-linear on-resistance [13], degrade the performance of a charge
redistribution SAR. The top plate switches, those that sample the input and those that per-
form DAC switching, are ideally sized with large aspect ratios for want of low on-resistance
and thereby higher sampling and DAC settling speeds. However, larger transistors worsen
the aforementioned errors, thus posing a trade-off between speed and precision.

2-2 Dynamic Performance Measures

Though INL and DNL give a fair measure of static linearity, it is useful to measure the ADC
performance in the frequency domain, which will include the linearity degradation at high

Master of Science Thesis Iniyavan Elumalai


14 Speed Limitations in SAR-based A/D Converters

frequencies and large signal amplitudes. The performance is usually determined by using a
single-tone sinusoidal input signal near the Nyquist frequency, and near full-scale amplitude.
Figure 2-5 shows a typical spectrum of the digital output from a 14-bit non-ideal ADC.
The input sinusoid shows up at the fundamental, and the white noise floor represents the
quantization noise (assuming thermal noise floor is superseded). The visible harmonics in the
spectrum are due to various non-linearities in the system; some of them may have been those
folded from higher Nyquist bands.

-20
Fundamental
Normalized Amplitude [dB]

-40

-60
Largest spur

-80

-100

0 fs /2
Frequency

Figure 2-5: Dynamic performance of a non-ideal 14-bit ADC

Signal-to-noise ratio (SNR) is defined as the ratio of the fundamental signal power to the
total noise power at the output, expressed as
 
Signal Power
SN R = 10log dB. (2-4)
Total noise power

The noise power represents the quantization noise and thermal noise within the Nyquist
bandwidth, excluding the harmonic components. It should be noted that the SNR is limited
only by quantization noise in the case of an input sinusoid, and can be written as

SN R = 6.02N + 1.76dB, (2-5)

where N is the ADC resolution.


Spurious free dynamic range (SFDR) is the ratio of the power at the fundamental to that
of the largest signal component in the spectrum within the Nyquist bandwidth. Signal-to-
noise and distortion ratio (SNDR) is the ratio of the fundamental signal power to the total

Iniyavan Elumalai Master of Science Thesis


2-3 Speed Limitations of a SAR ADC 15

noise and distortion power within the Nyquist bandwidth. SNDR is written as
 
Signal Power
SN DR = 10log10 , in dB. (2-6)
Total noise and distortion power
It is useful to measure SNDR as a function of amplitude, which will show that the quantization
and thermal noise dominates at low signal levels, whereas distortion dominates at higher signal
levels. When plotted as a function of frequency, it can be seen that SNDR degrades at high
frequencies due to effects such as clock jitter, high-frequency non-linearities of input sampling
network, comparator and reference buffer, etc.
Effective number of bits (ENOB) is a measure based on the SNDR of an ADC with a
full-scale sinusoidal signal near the Nyquist frequency. It is determined by
SN DR − 1.76
EN OB = . (2-7)
6.02

2-3 Speed Limitations of a SAR ADC

In a SAR conversion, the circuits are operated serially. After sampling the input, the converter
alternately repeats the operation of DAC settling and comparator resolving. In a conventional
synchronous ADC, an internal high-speed clock allocates every bit cycle to include time for
the digital logic, DAC settling and the comparator as shown in Figure 2-6. The time allocated
for each operation should be long enough so as to ensure proper operation in the worst cases.
The allocated time for DAC settling (t1 ) should cater to the worst-case settling requirement,
which is during MSB settling. The minimum allocated time for comparison (t2 ) has to ensure
that the comparator can resolve any input larger than 1/2 LSB.

t1
Comp. resolving
DAC settling
t2
Digital logic

Figure 2-6: Timing in a synchronous clocking scheme

It can be understood that for higher resolution converters, where more such bit cycles
have to be accommodated within a given clock speed, these allocated times for different
operations have to be reduced. Improving DAC settling, comparator resolution and digital
logic speed all imply power-hungry solutions, that may not be possible to realize at a given
technology node. This explains the absence of SAR converters at high conversion speeds.

2-4 Popular Techniques in Relevant Literature

SAR converters show excellent energy efficiency at moderate speeds. They are extensively
used in a time-interleaved architecture, which improves the speed of the total ADC by the in-

Master of Science Thesis Iniyavan Elumalai


16 Speed Limitations in SAR-based A/D Converters

terleaving factor. A big wave of current research efforts [14], [15], [16], [17] is working towards
resolving the errors caused by time-interleaving with minimal overhead. Improvements in the
speed of a SAR converter can also be expected from device scaling, which greatly enhances
the digital logic speed, and reduces the comparator resolving time. This section reviews a few
popular techniques reported in the recent literature that attempt to relieve the SAR converter
of its speed limitations at high resolutions.

2-4-1 Asynchronous Clocking

In a conventional, synchronous charge redistribution SAR ADC, as mentioned earlier, the


times allocated for different essential operations are fixed by their respective worst-case re-
quirements. It is to be noted that in a typical regenerative latch-based comparator (with
no loss of generality), that the resolving time has an inverse exponential dependence on the
voltage difference to be resolved [18]. Therefore, with a scheme based on worst-case resolving
time requirements, there are some situations when the comparator resolves very quickly, and
therefore does not use all of the time assigned to it. This wasted comparison time can be
eliminated.
Asynchronous clocking [19] employs a self-timed SAR loop. A fixed amount of time is
given for DAC settling, but not for comparator resolving. Whenever the comparator makes
a decision, it triggers a ready signal that initiates DAC settling for the next bit-cycle, as
illustrated in Figure 2-7. This way, the bit cycles vary dynamically thus eliminating the
wasted comparison time, and thereby improving the SAR speed.

Comp. resolving
ready
DAC settling

t2

Figure 2-7: Timing in an asynchronous clocking scheme

While asynchronous clocking saves time with regard to comparator resolving, the time
allocated for DAC settling (fixed by the worse-case settling requirement) is still an impediment
to improving the speed further. A semi-synchronous clocking scheme is proposed in [20] which
attempts to save more time by compromising on DAC settling appropriately. Figure 2-8
illustrates a SAR conversion which involves easy and hard decisions. A decision is easy when
the comparator can resolve quickly because the input voltage is far from the reference. On
the other hand, a hard decision is when metastability occurs because the input voltage is too
close to the reference, and the comparator takes a long time to resolve. It can be noted that
a hard decision is always followed by one or more easy decisions. In that case, DAC settling
can be relaxed for a cycle following a hard decision. Since an easy decision can be followed
by either hard or another easy decision, it is safe not to relax DAC settling for the next cycle.
It can be observed that in a bit cycle, if the comparator takes more time to resolve, DAC
settling can be given less time and vice versa. Given these observations, the semi-synchronous
scheme allocates equal times for all bit cycles, where the two operations - DAC settling and

Iniyavan Elumalai Master of Science Thesis


2-4 Popular Techniques in Relevant Literature 17

comparator resolving - have their times allocated dynamically . With shorter bit-cycles, this
scheme improves the entire SAR speed. For a deeper understanding of this scheme and its
implementation, the reader is directed to [20].

Vref

Vdac easy
Vref /2
easy
hard

hard

0 time

Figure 2-8: SAR conversion illustrating easy and hard decisions

2-4-2 Overranging

It is established from the previous sections that the DAC settling time allocated in a conven-
tional, binary SAR has to be long enough so that the settling error is less than 1/2 LSB in
every bit cycle. If the design is modified such that more settling error can be tolerated, the
converter does not have to wait for the DAC to settle completely and requires a shorter bit
cycle, thereby improving conversion speed. Building tolerance to a certain amount of DAC
settling error is possible by the principle of overranging, which is explained next.
In a binary search, the effective input range is reduced by a factor of two after every
DAC switching event. In such cases, no DAC settling error can be tolerated as exemplified
in Figure 2-4, where incomplete DAC settling led to an erroneous conversion. On the other
hand, if the search follows a non-binary radix, that is, the effective signal range reduces by
a factor less than two every bit cycle (as shown in Figure 2-9), a certain amount of settling
error can be tolerated within the allowable overrange.
In this case, more decision levels are generated and different digital codes can represent
the same input voltage. As Figure 2-9 illustrates, this gives an opportunity to fix the previous
decision error. This way of building in a redundancy, however, comes at the cost of having
to accommodate more bit cycles in order to reach the same accuracy as that of a binary
conversion. For example, assume a converter employing a search with radix 8/5, that is, the
effective signal range in every bit cycle being 5/8th of the previous range. There is 12.5%
overrange built in. In such a case, DAC settling error in the earlier cycles can be up to 12.5%

Master of Science Thesis Iniyavan Elumalai


18 Speed Limitations in SAR-based A/D Converters

Vref
15
14
13
12
11
10
9
Vdac 8
7
6
5
4
3
2
1
0
0
1 0 0 0 0

time

Overrange

Figure 2-9: Effect of incomplete DAC settling on the conversion result in the presence of
overranging

of the desired value without affecting the conversion result. Since log(8/5)/log 2 = 0.68 bit
is resolved each cycle, 21 bit cycles are required to reach 14 bits of accuracy.
Usually, a DAC with non-binary weighted capacitors enables the non-binary search radix
[12]. An arithmetical unit calculates the equivalent binary outputs from all different digital
codes (in this case, the unit maps the 221 comparison patterns to 214 equivalent outputs). With
overranging giving room for DAC settling error in the earlier bit cycles, the time allocated
for DAC settling per bit cycle can be reduced, thereby improving conversion time.

2-4-3 Pipelining

The popular technique of applying a pipeline to SAR ADCs combines the excellent energy
efficiency of SAR with the higher speed possibilities of a pipeline configuration [21], [22],
[23]. Figure 2-10 shows an ADC architecture where two SAR converters are pipelined for
higher speed. The coarse SAR ADC on the left generates the most significant bits. The
residue of this SAR conversion is the difference between the input signal and the final DAC
voltage. It is amplified by a residue amplifier and then fed into the fine SAR ADC on the
right, which generates the least significant bits. In this way, the second stage can be digitizing
the residue from the previous coarse conversion while the first stage is digitizing the input.
With both the stages processing different samples concurrently, the total ADC speed depends
only on the speed of each stage which, being a medium-resolution SAR, can operate at high

Iniyavan Elumalai Master of Science Thesis


2-5 Summary 19

conversion speeds. This architecture offers several other advantages over a conventional SAR-
only architecture besides higher throughput, which will be discussed later in Chapter 5.

Stage 1 Stage 2

Vres
RA
Vref 1 Vi Vref 2

DAC DAC

SAR Logic SAR Logic

Dout
Figure 2-10: A two-stage, SAR-assisted pipeline ADC

2-5 Summary

In this chapter, several basic non idealities associated with a charge redistribution SAR ADC
are discussed. The fundamental speed limitations in a SAR ADC are identified as the stringent
requirements on DAC settling and comparator resolving at high resolutions, which can be
overcome at the expense of power-hungry solutions. Finally, we looked at some promising
techniques in the recent literature that attempt to relieve the SAR of its speed limitations.

Master of Science Thesis Iniyavan Elumalai


20 Speed Limitations in SAR-based A/D Converters

Iniyavan Elumalai Master of Science Thesis


Chapter 3

Proposed Reference Scheme

It has been mentioned earlier in Chapter 1 that creating different DAC reference levels by
charge redistribution causes code-dependent errors on the reference voltage during SAR con-
version. In this chapter, we look at how the conventional techniques that deal with these
errors limit the SAR speed in the case of high-resolution converters with given power and
area budgets. Following this, the reference scheme proposed in this research project is pre-
sented. Finally, the reference requirements for the prototype ADC are mentioned briefly.

3-1 Conventional Reference Circuitry

In a charge redistribution SAR, the capacitive DAC creates the required reference levels by
switching different amounts of capacitance out of the total array capacitance Cdac to/from
the reference source every bit cycle. Figure 3-1 shows a simplified representation of the
DAC during SAR conversion, where the reference circuitry comprising a buffer and bypass
capacitance Cref is loaded dynamically, with α varying every bit cycle as dictated by the
switching decisions.

Vref αCdac
Reference Vdac
from bandgap
Cref (1 − α)Cdac

Code-dependent network

Figure 3-1: Code-dependent loading of reference source

Master of Science Thesis 21 Iniyavan Elumalai


22 Proposed Reference Scheme

Such dynamic loading draws code-dependent charge from the reference circuitry. Since
the buffer is unable to supply this charge instantaneously on account of its finite output
impedance, the voltage Vref suffers instantaneous code-dependent drops, from which the
buffer recovers over time. Figure 3-2 shows the different possible drops on the reference
voltage after every DAC switching event in a 5-bit differential set-and-down switching scheme
(which is explained in the next section), and illustrates how they are recovered by a linear
settling buffer in every bit cycle before comparison.

Convert
Sample #4 #3 #2 #1 #0 Sample

Vref

∆3

time

Figure 3-2: Code-dependent transients on Vref recovered by the charge supplied by the
buffer

After an ith DAC switching event (i is the bit cycle index that takes the values
{N − 1, N − 2, ..., 1, 0} from MSB to LSB in a SAR Convert phase) has caused a drop of
∆i = Qi /Cref in the reference voltage, where Qi represents the corresponding code-dependent
charge drawn from reference, the buffer recovers as given by
Vref (t) = Vref − ∆i e−t/τ . (3-1)
Here,
τ = Rout (Cref + Cdac ) (3-2)
is the settling time constant of the buffer, and Rout , its output impedance. For a binary SAR
ADC with no overrange, the settling error on Vref before every comparison has to be less
than 1/2 LSB [12]. Thus, in every bit cycle, the error in Equation (3-1) should follow the
condition:
1 VF S,dif f
∆i e−tc /τ < volts, (3-3)
2 2N
where tc is the time instant at which the comparator is fired, which can be, for example,
halfway through the bit cycle. VF S,dif f is the differential, full-scale range of the ADC, and
N is the ADC resolution. Enough DAC settling time should be given so that this condition
is satisfied during every bit cycle. In fact, it is usually determined by the worst-case settling
requirement after MSB switching, which involves switching the largest capacitance that causes
the largest drop in reference voltage. In order to reduce the time allocated for DAC settling
and thus improve the conversion speed at a given resolution, the output impedance of the
buffer has to be lowered. The output impedance for a simple source follower is given by 1/gm ,
which can be reduced only by increasing the static bias current. Circuit techniques such as
class-AB operation and shunt feedback are usually employed in order to obtain low output
impedance without requiring large static bias currents.

Iniyavan Elumalai Master of Science Thesis


3-1 Conventional Reference Circuitry 23

As the resolution required of the converter increases to above 12 bits, designing such a
high speed buffer may not be feasible within a given power budget. This is because the settling
requirements on reference voltage, stated by Equation (3-3), grow stringent in multiple ways:
(1) the large array capacitance (for the sake of high resolution) demands large reference charge
during DAC switching, which worsens the drops (∆i ) in the reference voltage. (2) the value
V
of an LSB (= F S,dif
2N
f
) becomes smaller for the same full-scale range. (3) more bit cycles
have to be accommodated. In such a case, reducing the output impedance to a low enough
value becomes impossible within a given static power budget for the reference buffer, which
requires the time allocated for DAC settling to be increased. This essentially limits the whole
ADC conversion speed.
For example, for a 14-bit differential converter having a full-scale range of 2Vref with
differential reference voltages Vref p = Vref and Vref n = 0 V, satisfying Equation (3-3) in the
worst case requires
1 2Vref
∆wce−tc /τ < volts. (3-4)
2 214
The worst-case drop in the reference voltage (∆wc ) happens when the MSB capacitor (Cdac /2)
is switched to halve the reference voltage during the first step in the binary search (Refer
Figure 1-4(a)). The reference charge in that case is given by
  
Cdac Vref
Qwc = , (3-5)
2 2
which causes the worst-case drop of
Cdac
∆wc = Vref . (3-6)
4Cref
Thus, Equation (3-4) becomes
Cdac −tc /τ 1
e < 14 V. (3-7)
4Cref 2
Assuming Cdac = Cref ,
e−tc /τ < 2−12 V. (3-8)
Thus, 8.32τ of DAC settling is required within each bit cycle before comparison at tc . For
the converter to operate at 100 MHz sampling speed with 50% duty-cycle between Sample
and Convert phases, each bit cycle can be budgeted as
1 1 1
6
× × s = 300 ps + 8.32τ, (3-9)
100 × 10 2 14
where 300 ps is allocated for other operations in a bit cycle, namely, comparison and digital
logic. This is a reasonable assumption for the 28 nm CMOS technology node. This budgeting
yields the required time constant τ = 6.868 ps. In order to calculate the required output
impedance of the buffer from Equation (3-2), an assumption on Cdac can be made based on
the thermal noise budget for the converter. For a 12-bit thermal noise performance (i.e.,
74 dB SNR according to Equation (2-5)) with 1.4 Vpp,dif f signal range, the SNR of the
converter can be written as,
 
1.4

2 2
SN R = 20log  q  = 74dB, (3-10)
2kT
Cdac

Master of Science Thesis Iniyavan Elumalai


24 Proposed Reference Scheme

q
where C2kTdac
represents the total sampled noise voltage on a differential DAC. This equation
yields the single-ended array capacitance Cdac = 1.4 pF. Thus, with τ = 6.868 ps, and
Cdac = Cref = 1.4 pF, the required output impedance of the buffer has to be less than 3 Ω.
With a simple source follower, this demands Id = gm2Vod = 150mV
2×3 = 25 mA, which from a
1.8 V supply will represent a major fraction of the entire ADC current consumption.
It should be noted that, in the calculation above, the bypass capacitance Cref was as-
sumed to be equal to the array capacitance, Cdac . However, the choice of Cref requires a
careful analysis. Rewriting Equation (3-7) yields
  
Cref
tc > τ 8.32 − ln . (3-11)
Cdac

This represents the DAC settling time to be allocated every bit cycle. Substituting for τ from
Equation (3-2) gives
"   #
Cref
DAC settling time = Rout (Cref + Cdac ) 8.32 − ln . (3-12)
| {z } Cdac
τ | {z }
number of time constants

C
This equation shows that for a given Rout , reducing the ratio of Cref dac
decreases the time
constant and increases the number of time constants required, the former by a greater extent.
Figure 3-3 shows a surface plot of the current required in the reference buffer versus the ratio

140
Current required in the reference buffer [mA]

120

100

80

60

40

20
150
0 100
10 8 6 4 2 50 Sampling clock
0 frequency [MHz]
Cref/Cdac

Figure 3-3: Surface plot of the current required in the reference buffer versus sampling
Cref
frequency and the ratio Cdac

C
( Cref
dac
), and the sampling clock frequency fclk for the aforementioned converter. As the plot
indicates, reducing Cref is beneficial. Thus, it is preferable to leave the buffer’s output node
agile, with only the parasitic capacitances present there comprising Cref [13]. Nevertheless,

Iniyavan Elumalai Master of Science Thesis


3-2 A Stand-alone Capacitor as the Reference Buffer 25

the strong coupling between the power consumption of the buffer and the conversion speed is
evident in this approach.
Another conventional technique is to increase the bypass capacitance Cref to a value
such that switching-induced disturbances from the DAC on the reference voltage are reduced
to less than 1/2 LSB. With this approach, Equation (3-7) requires Cref to be equal to 212
times Cdac , or approximately 6 nF, which will dominate the chip area or requires an off-chip
approach which would limit the speed.

3-2 A Stand-alone Capacitor as the Reference Buffer

From the section above, it becomes clear that the conventional techniques of dealing with the
switching-induced errors on the reference voltage limit the energy efficiency at higher speeds.
Thus, for high-end converters, a power-conscious scheme to provide reference voltages has to
be devised. The conventional active buffer essentially burns power to provide a low output
impedance (or a high slew rate) that enables quick recovery of the reference voltage from the
switching-induced errors up to the required precision. It should be recalled from Section 2-4
that overranging helps relax the settling precision required, thus promising power savings.
However, it also requires more bit cycles in order to obtain the same accuracy as that of a
binary conversion. Thus, the power-speed trade-off is not completely relaxed.
In that regard, this research project proposes to employ a stand-alone capacitor in place
of the conventional buffer. The capacitor tied to the reference buffer is disconnected during
SAR conversion, thus preventing reference recovery and thereby eliminating the speed limit
set by the output impedance of the buffer. In order to reduce the switching-induced errors
on the reference voltage to less than 1/2 LSB, the capacitance has to be as large as ∼
2N times the array capacitance, as discussed earlier. However, the scheme proposed here
employs a capacitance only 10 - 20 times the array capacitance, and deals with the resulting
code-dependent errors through calibration schemes (discussed in Chapter 4). Figure 3-4(a)
illustrates the proposed reference buffer scheme. The reference capacitor remains tied to the
buffer during the Sample phase of the ADC. As the SAR conversion starts, the capacitor is
disconnected from the buffer and stands alone, holding the reference voltage. DAC switching
pulls code-dependent charge from the stand-alone capacitor. Since its capacitance is not
infinitely high, the reference voltage suffers significant code-dependent errors. Figure 3-4(b)
shows the resulting reference voltage with this scheme for all the possible codes as those
in Figure 3-2. In the next Sample phase, the capacitor is replenished by the buffer, which
removes the code-dependent error in the reference voltage at the end of SAR conversion up
until the precision required of the ADC.
With this scheme, the disturbances on the reference voltage are no longer recovered by
the buffer during the Convert phase and become static errors. They can be formulated,
as illustrated in the next section, given the knowledge of how the DAC is being switched.
Thus, as long as these errors can be handled by the ADC, they can be corrected through
calibration schemes that need not be power-hungry. The ADC may not handle the dropping
reference voltage due to various reasons. For example, the decreasing reference voltage causes
the common mode voltage at the comparator to vary largely, which the comparator may not
handle. Also, in the case of a pipelined-SAR converter, if the reference voltage in the first

Master of Science Thesis Iniyavan Elumalai


26 Proposed Reference Scheme

Convert
Vref
Reference
from bandgap
Cref

(a)

Convert
Sample #4 #3 #2 #1 #0 Sample

Vref

∆3

time

(b)
Figure 3-4: (a) Reference capacitor disconnected during SAR conversion (b) Code-
dependent, deterministic errors on Vref

stage drops considerably, the residue voltage at the end of the conversion may exceed the
range that can be handled by the ensuing back-end. This essentially sets the lower limit on
the value of the reference capacitor relative to the array capacitance, which determines the
magnitude of the drops in the reference. It should be noted that the lower limit on Cref can
also be set by the reference noise, which in this case is simply kT /Cref .

3-3 Deterministic Errors on the Reference Voltage

Before discussing the benefits that the proposed reference scheme entails, this section proves
that the switching-induced errors on the reference voltage held by a stand-alone capacitor are
deterministic, which enables them to be corrected by calibration schemes. Formulating these
code-dependent errors requires knowledge of how the DAC is being switched to create the
required reference levels from the reference source. There are several switching algorithms
available in the literature [24, 25, 26, 27, 28, 29], one of which is set-and-down switching [26].

Iniyavan Elumalai Master of Science Thesis


3-3 Deterministic Errors on the Reference Voltage 27

Vref
2C C C

Vip
Vi > 3Vref /4?
Vref Vref Vin
2C C C Yes
2C C C
Vip
Vi > Vref /2?Vref Vref Vref
Vin
Vref Vref
No 2C C C
2C C C
Vref Vref Vref Vip
Vref Vref Vref Vref Vref Vref Vi > Vref /4?
Vin
2C C C 2C C C Yes
Vip Vip 2C C C
Vi > 0? Vref Vref
Vin Vin
Vref Vref
2C C C
2C C C 2C C C
Vref Vref Vref No
Vref Vref Vref Vip
Vi > −Vref /4?
Vref Vref Vref Vin
2C C C
Yes 2C C C
Vip Vref Vref
Vi > −Vref /2?
Vin Vref Vref Vref
2C C C
No
2C C C
Vref Vref Vip
Vi > −3Vref /4?
Vin

2C C C
Vref

Sample i=2 i=1 i=0 time


Figure 3-5: 3-bit set-and-down switching procedure

Figure 3-5 illustrates a typical set-and-down switching procedure for a 3-bit differential
DAC. During the Sample phase, the differential capacitive DAC (two arrays, each comprising
{2C, 1C, 1C}, where C is the unit capacitance) samples the input voltage Vi on their top plates
that are all connected to the input of the comparator. Meanwhile, the bottom plates of all
the capacitors are connected to the positive reference voltage Vref , as shown in the leftmost
part of the figure. After the input charge is thus trapped in the array during the Sample
phase, the top plates of the array capacitors are disconnected from the input, and the SAR
conversion starts. No switching is required during the MSB bit cycle (i = 2), as the input
voltage is already present at the input of the comparator directly enabling the comparison
(Vi > 0?). Based on the answer to this question, i.e., comparator decision, the next bit-cycle
(i = 1) involves one of these comparisons: (Vi > Vref /2?) and (Vi > −Vref /2?). For the
former, the bottom plate of the MSB capacitor in the positive (upper) DAC is disconnected
from Vref and connected to the negative reference (assumed to be ground). For the latter,
the bottom plate of the MSB capacitor in the negative DAC is switched from Vref to ground.
The comparator decision from this bit cycle decides how the switching is done in the next,
and this continues until the LSB bit cycle (i = 0). The capacitors being switched during
every bit cycle are highlighted (in red) in the figure.

Given the switching algorithm, it becomes possible to determine the decision-dependent


charge pulled out of the reference source in every bit cycle. For example, in the illustration
above, zero charge is required of the reference during the MSB bit cycle as there is no switching
involved. Whereas in the next bit cycle i = 1, there are two cases of switching. It can be
noticed that both the cases can be represented by a single switching scenario, as shown in
Figure 3-6. The MSB capacitor (2C) is switched from Vref to ground, changing the DAC

Master of Science Thesis Iniyavan Elumalai


28 Proposed Reference Scheme

voltage from Vdac = Vref to Vdac′ = Vref /2. The charge supplied by the reference source
during this switching is given by

Q1 = −(C + C)(Vdac − Vdac )
 
Vref
= (2C) = CVref . (3-13)
2

Vref Vref

2C C C Q1 = CVref C C

′ = Vref
Vdac = Vref Vdac 2
2C

Figure 3-6: Switching scenario in bit cycle i = 1

Similarly, in the next bit cycle i = 0, there are four cases of switching, which can be
grouped under two different switching scenarios as shown in Figure 3-7.

Vref Vref
CVref
C C Q0,a = 4 C

Vref ′ = Vref
Vdac = 2
Vdac 4
2C 2C C

(a) Switching scenario for D2 D1 = 00 or 11

Vref Vref
3CVref
2C C C Q0,b = 4 2C C

′ = 3Vref
Vdac = Vref Vdac 4
C

(b) Switching scenario for D2 D1 = 01 or 10

Figure 3-7: Two different switching scenarios in bit cycle i = 0

Iniyavan Elumalai Master of Science Thesis


3-3 Deterministic Errors on the Reference Voltage 29

In Figure 3-7(a) illustrating one of the scenarios, the MSB - 1 capacitor (C) is switched
in order to change the DAC voltage from Vdac = Vref /2 to Vdac ′ = Vref /4. The reference
charge in this case is given by
 
Vref Vref CVref
Q0,a = −(C) − = . (3-14)
4 2 4

For the other scenario depicted in Figure 3-7(b), the switching is performed to change the
′ = 3V
DAC voltage from Vdac = Vref to Vdac ref /4. The corresponding reference charge is given
by  
3Vref 3CVref
Q0,b = −(2C + C) − Vref = . (3-15)
4 4

Q2 Q1 Q0
CVref
4

CVref

3CVref
0 4

3CVref
4

CVref

CVref
4

i=2 i=1 i=0 time


Figure 3-8: Different possible reference charges for a 3-bit set-and-down switching procedure

Figure 3-8 shows all of these code-dependent charges supplied by the reference in a sim-
plified representation of the DAC switching tree. A more generalized derivation of reference
charge for any given switching algorithm is provided in Appendix A. For an N-bit set-and-
down switching procedure, it is derived that the charge required in each bit cycle is given
by
!
i
P P
Ci+1 Cl + s
l=0 s∈Sd
Qi = (Vref p − Vref n ) (3-16)
Cdac
for i = N − 2, N − 3, ..., 0.

Sd ⊂ {Cm |(i + 1) < m < N }, the choice of subset being code-dependent. Thus, after every ith
switching event there are 2N −i−2 possibilities of reference charge based on switching decisions.
Figure 3-9 shows all of the possible switching reference charge values for an 8-bit set-and-down
switching algorithm. The code-dependent spread in reference charge reduces over time from
MSB to LSB.

Master of Science Thesis Iniyavan Elumalai


30 Proposed Reference Scheme

35

Qi (Reference charge per step) [CVref]


30

25

20

15

10

0
7 6 5 4 3 2 1 0
i (Bit cycle index)

Figure 3-9: Possible values of reference charge per step for an 8-bit set-and-down switching
procedure

75
Total reference charge [CVref]

70

65

60

55

50

45

40
0 50 100 150 200 250
Output Code

Figure 3-10: Total reference charge versus output code for an 8-bit set-and-down switching
procedure

The total charge supplied by the reference source during the Convert phase is related the
dynamic power dissipated by the reference circuitry. Figure 3-10 shows the total reference
charge versus output code for the 8-bit case. In order to reduce this, the value of the unit
capacitance can be reduced to the limits allowed by kT /C noise and matching requirements.
Moreover, several switching algorithms such as split-capacitor [24], energy-saving [25], merged

Iniyavan Elumalai Master of Science Thesis


3-4 Benefits of the Proposed Reference Scheme 31

capacitor [27], input-range prediction [28], and charge-average switching [29] are proposed in
the literature in attempts to save dynamic power. However, this dynamic power is only a small
fraction of the total power dissipated by the reference circuitry, especially at high conversion
speeds [8], [9]. As discussed earlier in Section 3-1, the main cause of power consumption in
the reference circuitry is that the buffer must provide a low output impedance in order to
settle the code-dependent errors on the reference voltage within every bit cycle.

3-4 Benefits of the Proposed Reference Scheme

The proposed reference scheme, in a bid to eliminate the energy limit set by the buffer,
employs a stand-alone capacitor to hold the reference voltage during SAR conversion. As the
employed capacitance is only ∼ 10 - 20 times the DAC array capacitance, DAC switching
causes code-dependent errors on the reference voltage. The previous section illustrated that
those static errors on the reference voltage can be predicted, and thus can be corrected through
calibration schemes.
With this proposed reference scheme, the DAC settling speed and output impedance of
the reference buffer have been decoupled to a certain extent. Since the DAC settling in this
case is limited only by the on-resistance of the switches connecting the reference capacitor
to the capacitive DAC, the bit cycle times can be reduced significantly, thus improving ADC
conversion speed. This may not bring power savings directly, as the buffer still has to replenish
the reference capacitor during the Sample phase (shown in Figure 3-3). However, it can be
noted that in the conventional scheme the buffer has to recover the reference voltage in short
intervals, which demands large currents, especially in case of a high resolution converter.
With the proposed scheme, the buffer can recover the voltage during a longer interval - i.e.,
the whole Sample phase - and thus power savings can be expected. Nonetheless, this scheme
improves the ADC speed without compromising the power efficiency.

3-5 Design Requirements

The proposed reference scheme is implemented in the two-stage SAR-assisted pipeline ADC
shown in Figure 3-11(a). The effective performance of an ADC is usually limited by ther-
mal noise, quantization noise, and harmonic distortion, of which thermal noise essentially
determines the power dissipated in case of high resolution (> 12-bit) converters. This ADC,
aiming for a 12-bit ENOB performance at 50 MHz sampling clock, is thus designed such that
its effective resolution is limited only by thermal noise. In order for the quantization noise
not to limit the performance, the ADC quantizes the input at a 14-bit level.
With an input full-scale range of 1.4 Vpp,diff , the first stage in the pipeline resolves 7 bits
resulting in a residue voltage in the range 10.9375 mVpp,diff . The residue, after being ampli-
fied 16x, is passed on to the second stage. This residue amplification reduces the precision
requirements on the second stage by 16x. It is to be noted that the input linear range of
this amplifier basically determines the first stage resolution. The second stage, designed for
an input range of 375 mVpp,diff resolves 8 bits. Doing so effectively introduces 2x overrange

Master of Science Thesis Iniyavan Elumalai


32 Proposed Reference Scheme

between the two stages, which is beneficial since comparator noise in first stage can be toler-
ated with the additional ∼ 10 mVpp,diff overrange thus introduced at the input of the residue
amplifier. This overrange also helps the ADC tolerate up to 2x error in the interstage gain.
The timing diagram is shown in Figure 3-11(b). The two stages in the pipeline operate
in a complementary fashion. During their respective Sample phases, the first stage samples
the input, and the second stage samples the residue of the previous conversion from the first
stage. While the Convert phase for the second stage is completely dedicated to its 8-bit SAR
conversion, the first stage employs ∼ 1/10-th of its Convert phase for residue amplification.

Cs = Cdac + Catt = (1 + 0.4) pF Cs = (100 + 60) fF


16x
10 mVppd
RA
αCdac (1 − α)Cdac Catt αCdac (1 − α)Cdac Catt

1.4 Vppd 375 mVppd


1.0 V 7b 0.8 V 8b
0V 0.5 V

(a) Block diagram

RA

Stage 1 Sample Convert

Stage 2 Convert Sample

5 ns 4 ns 1 ns

(b) Timing diagram

Figure 3-11: Two-stage, SAR-assisted pipeline ADC

As for the reference voltages, the first stage employs Vref p = 0.7 V and Vref n = 0 V in
order to achieve 1.4 Vpp,diff input range, the 0.7 V required being achieved by attenuating
1 V. This reference attenuation is done by connecting a fraction of the total sampling DAC
capacitance to ground during Convert phase. In this case, 0.4 pF out of the total 1.4 pF sam-
pling capacitance Cs (designed for 12-bit kT /C noise) is connected to ground. This grounded
capacitance helps reduce the comparator kickbacks during the Convert phase. Similarly, the
second stage employs attenuated voltages from 0.8 V and 0.5 V for positive and negative
reference voltages, respectively. Reference attenuation in this stage is performed with 100 fF
out of the total 160 fF sampling capacitance, thereby covering a signal range of 375 mVpp,diff .
With the total ADC operating at the 12-bit thermal noise level, noise on first stage
reference is budgeted at 12% of the total noise power, or ∼ 37 uV rms, assuming a uniformly
distributed output code (recall that the reference noise is multiplicative with output code).
The noise on the second stage reference is budgeted at around 5%, or ∼ 25 uV rms. The
precision requirements on the reference voltages for each stage are determined by the amount

Iniyavan Elumalai Master of Science Thesis


3-6 Summary 33

of overrange built into their respective capacitive DACs. More specific details will be given
in the course of the next two chapters.

3-6 Summary

In this chapter, we looked at conventional techniques of providing a steady reference to the


converter. Identifying that such techniques lead us to the point of diminishing returns in
the case of high-end converters, this chapter presented a new reference buffer scheme that
alleviates the strong coupling between ADC speed and static power of the reference buffer.

Master of Science Thesis Iniyavan Elumalai


34 Proposed Reference Scheme

Iniyavan Elumalai Master of Science Thesis


Chapter 4

Charge-based Compensation of
Reference Errors

In this chapter, we investigate the nonlinearity in the ADC transfer caused by code-dependent
errors on the reference voltage held by a stand-alone capacitor. Following this, a simple tech-
nique of dealing with this reference-induced distortion is discussed. Finally, the implementa-
tion details of the proposed scheme in the fine stage of the prototype two-stage, SAR-assisted
pipeline ADC are given.

4-1 Reference-induced Distortion

The code-dependent errors induced on the reference voltage due to DAC switching are signifi-
cant with a stand-alone reference capacitor valued Cref that is not much larger than the DAC
array capacitance, Cdac . Figure 4-1 shows all of the possible code-dependent errors on Vref
= 1 V in the case of an 8-bit, differential, set-and-down DAC with Cref = Cdac (simulated
in MATLAB). When plotted against the code available at the end of every bit cycle, the
total drop in the reference voltage until then shows a parabolic pattern. This is illustrated in
Figure 4-2, where the reference voltage at the end of the last bit cycle is plotted against the
complete 8-bit code. A quadratic dependence of reference voltage on the digital output code
translates to an odd-order non-linearity in the DAC transfer function, and thus the whole
ADC transfer. This is because, as mentioned in Section 2-1-4, any error on the reference
has a signal-dependent gain towards the output. Thus, quadratic error on the reference (in
this case), when multiplied by the signal-dependent gain, leads to odd-order distortion in the
ADC.
Figure 4-3 shows the 4096-point FFT output for an actual 8-bit ADC (with a binary-
weighted DAC array following set-and-down switching algorithm) sampling a 90% FS,
3.1 MHz input signal at 100 MS/s. Using Cref = Cdac leads to odd-order distortion com-
ponents well above the 8-bit quantization noise floor that is represented by the dashed red
line at 50 dB. With the 3rd harmonic representing the largest spur at 22.1 dB below the

Master of Science Thesis 35 Iniyavan Elumalai


36 Charge-based Compensation of Reference Errors

0.95

0.9

0.85

0.8
Vref

0.75

0.7

0.65

0.6

0.55
7 6 5 4 3 2 1 0
i (Bit cycle index)

Figure 4-1: Code-dependent errors on Vref = 1 V in the case of 8-bit set-and-down switching
with Cref = Cdac (simulated in MATLAB)

0.76

0.74

0.72

0.7

0.68
end
Vref

0.66

0.64

0.62

0.6

0.58
0 50 100 150 200 250
8−bit output code

Figure 4-2: Reference voltage at the end of the Convert phase versus 8-bit output code
(simulated in MATLAB)

fundamental, the total harmonic distortion (THD) is equal to 20.1 dB which essentially sets
the SNDR. This represents an ENOB of a mere 3.04 bits. It should be noted that if the spec-
trum of the digital output code were obtained along with the conversion residue, the total
reference-induced distortion would show up as a single, large, third-order harmonic because
of the smooth quadratic dependence of Vref on the output code (shown later in Chapter 5).

Iniyavan Elumalai Master of Science Thesis


4-1 Reference-induced Distortion 37

−10 SFDR = 22.1 dB


SNDR = 20.1 dB
−20 ENOB = 3.04 bits

−30
Amplitude [dB]

−40
Quantization noise floor
−50

−60
FFT Gain = 10 log(4096/2)
−70

−80 FFT noise floor

−90

−100
0 10 20 30 40 50
Frequency [MHz]

Figure 4-3: Performance of an actual 8-bit ADC with binary-weighted DAC and stand-alone
reference capacitor, Cref = Cdac

Increasing the ratio of Cref /Cdac reduces this reference-induced distortion directly, as
evident in Figure 4-4, which shows the FFT output for the same input signal in the case
of Cref = 10Cdac . Values of SFDR, THD and SNDR are 42.3 dB, 38.9 dB and 38.5 dB,
respectively. A 10x increase in the value of Cref thus enhances the THD by 18.8 dB. In
order for the reference-induced distortion components not to degrade the total SNDR beyond
thermal or quantization noise limits, Cref has to be as large as 26 times Cdac (Section 3-1).

−10
SNDR = 38.5 dB
−20 ENOB = 6.1 bits
SFDR = 42.3 dB
−30
Amplitude [dB]

−40
Quantization noise floor
−50

−60
FFT Gain = 10 log(4096/2)
−70

−80 FFT noise floor

−90

−100
0 10 20 30 40 50
Frequency [MHz]

Figure 4-4: Performance of an actual 8-bit ADC with binary-weighted DAC and stand-alone
reference capacitor, Cref = 10Cdac

Master of Science Thesis Iniyavan Elumalai


38 Charge-based Compensation of Reference Errors

4-2 Error Compensation in Charge Domain

This section introduces a straightforward technique for correcting the switching-induced errors
on Vref held by a stand-alone capacitor, without having to increase its capacitance Cref
to large values and pay a penalty in terms of area. The charge pulled from the reference
capacitor after a DAC switching event can be predicted, given the switching decision and the
knowledge of the algorithm in use, as illustrated in Section 3-3. Knowledge of the charge lost
in a DAC switching event A posteriori enables it to be compensated for precisely. As soon as a
comparator decision initiates a switching event that pulls decision-dependent charge from the
reference capacitor, a corresponding compensation charge can be added to it instantaneously,
thus preventing Vref from suffering a drop. To that end, this technique employs a bank of
capacitors that have been pre-charged before SAR conversion, from which a certain amount
of capacitance is connected to the main reference capacitor after a DAC switching event. The
amount of capacitance chosen holds the charge required for compensation.

Vref Vrefi,α
-Qi,α

Cref Cref

(a) Losing charge from Cref

Vc

Vrefi,α Vref

Cref Cci,α Cref + Cci,α

(b) Compensating for the charge lost

Figure 4-5: Conceptual illustration of error compensation in charge domain

Figure 4-5 illustrates this technique in charge domain. After a DAC switching event i
has pulled Qi,α amount of charge from Cref , where α represents the code dependence, Vref
drops to Vrefi,α . From the compensation bank consisting of capacitors that have all been
pre-charged to a voltage Vc , Cci,α is chosen and connected to Cref . Instantaneous charge
sharing between the capacitors is such that the drop in Vref is compensated for. Though the
figure shows successively timed events of charge being lost, followed by compensation, both
the events can be simultaneous so that Vref does not suffer instantaneous drops.

Iniyavan Elumalai Master of Science Thesis


4-2 Error Compensation in Charge Domain 39

The compensation capacitor Cci,α should be sized according to the charge-conservation


equation given by  
Cref Vref − Qi,α + Cci,α Vc = Cref + Cci,α Vref . (4-1)
Since there are 2N −i−2 code-dependent possibilities of reference charge in every bit cycle
from (MSB - 1) to LSB (with i taking values {N − 2, N − 1, ..., 1, 0}) for an N-bit set-
and-down switching DAC, the bank is required to accommodate (2N −1 − 1) capacitors with
unique values. Since this would imply a huge penalty in terms of area, the compensation
bank employs small capacitances that are charged to Vc (i.e., higher than Vref ). Assuming
Cci,α << Cref , Equation 4-1 becomes
Cci,α Vc ≈ Qi,α . (4-2)
For an 8-bit DAC following set-and-down switching algorithm, Qi,α can range from 32Cu Vref
to 0.0078Cu Vref , where Cu represents the unit capacitance of the DAC. Thus, by choosing
Vc appropriately higher than Vref , compensation capacitors can have values less than or
approximately equal to the unit capacitance.
An example of this technique is discussed next. For a 5-bit set-and-down DAC, the
different possible drops in reference voltage are as shown in Figure 4-6. At t3 , the DAC
switching is at the start of bit cycle i = 3, corresponding to either of the previous decisions:
D4 = 0 and D4 = 1. It pulls a code-independent amount of charge Q3 = 4Cu Vref from Cref .
Vref drops to Vref3 . At t2 , the DAC switching in bit cycle i = 2 causes a code-dependent
drop from Vref3 . For D4 D3 = 00 and D4 D3 = 11, the amount of charge pulled from Cref is
Q2,a = Cu Vref3 . For D4 D3 = 01 and D4 D3 = 10, it is Q2,b = 3Cu Vref3 . Accordingly, Vref3
drops to either Vref2,a or Vref2,b . The reference voltage suffers such code-dependent drops as
the SAR conversion progresses.

Vref
Q3 = 4Cu Vref

Vref3
Vref2,a Q2,a = Cu Vref3
Q2,b = 3Cu Vref3
Vref2,b

t3 t2 t1 t0 time
i=4 i=3 i=2 i=1 i=0

Figure 4-6: Code-dependent errors on Vref in the case of 5-bit set-and-down switching

In order to compensate for these errors on Vref , a bank of compensation capacitors is


initially pre-charged to a voltage Vc . The first code-independent drop at the start of bit cycle
i = 3 is compensated for by connecting the capacitor Cc3 from the bank to Cref at t3 , as
shown in Figure 4-7. The value of Cc3 is given by
Q3 4Cu Vref
Cc3 = = , (4-3)
Vc Vc

Master of Science Thesis Iniyavan Elumalai


40 Charge-based Compensation of Reference Errors

Vref
Vc
Cref
Cc3

D4 + D̄4 Compensation bank


(a) Schematic illustration

Vref
Q2,a = Cu Vref
Q2,b = 3Cu Vref

t3 t2 t1 t0 time
i=4 i=3 i=2 i=1 i=0
(b) Vref after compensation at t3

Figure 4-7: Illustration of compensation in bit cycle i = 3

As Vref3 recovers to Vref , the code-dependent values of charge pulled at the start of bit cycle
i = 2 become Q2,a = Cu Vref and Q2,b = 3Cu Vref . In order to compensate for the former,
Cc2,a given by
Q2,a Cu Vref
Cc2,a = = (4-4)
Vc Vc
is chosen from the bank and connected to Cref at t2 . For the latter case, Cc2,b , given by

Q2,b 3Cu Vref


Cc2,b = = (4-5)
Vc Vc
is chosen. Both cases are shown in Figure 4-8. Similarly, compensation can be performed
in every bit cycle by choosing an uniquely-valued capacitor from the compensation bank for
each and every comparator decision. The digital logic performing the selection from the bank
should be as fast as the logic triggering the DAC switching event based on the comparator
decision. This prevents the reference voltage from suffering instantaneous drops. Figure 4-9
shows the reference compensation loop (in grey) in action alongside the main SAR loop.

Iniyavan Elumalai Master of Science Thesis


4-2 Error Compensation in Charge Domain 41

Vref
Vc
Cref
Cc2,a Cc2,b

(D4 D3 + D̄4 D̄3 ) (D4 D̄3 + D̄4 D3 ) Compensation bank


(a) Schematic illustration

Vref

t3 t2 t1 t0 time
i=4 i=3 i=2 i=1 i=0
(b) Vref after compensation at t2

Figure 4-8: Illustration of compensation in bit cycle i = 2

Vref

Cref Selection
Logic

dac_data Vdac comp_out


DAC

SAR Logic

Dout

Figure 4-9: Conceptual illustration of reference compensation loop (in grey) along with the
main SAR loop

Master of Science Thesis Iniyavan Elumalai


42 Charge-based Compensation of Reference Errors

4-3 Implementation Choices

This section examines some of the design choices concerning implementation of the technique
discussed above, and proposes the following improvements:

4-3-1 Systematically Decreasing Vref

The error in Vref per bit cycle can be considered to consist of two parts: the code-independent
drop and the code-dependent spread. It is the latter that causes distortion in the ADC. If not
for the code-dependent spread, Vref would only decrease systematically every bit cycle. This
is shown in Figure 4-10(b), where only the code-dependent spread is removed per bit cycle,
instead of correcting the total error in Vref (shown in Figure 4-10(a)). As evident, removing
the spread requires less compensation charge than that required when correcting the entire
error. Having smaller compensation capacitors is beneficial when chip area is considered.

Vref

t1 t2 t3 t4 time
i=3 i=2 i=1 i=0

(a)

Vref

t1 t2 t3 t4 time
i=3 i=2 i=1 i=0

(b)

Figure 4-10: Compensation for (a) constant Vref , and (b) code-independent, systematically
decreasing Vref

Iniyavan Elumalai Master of Science Thesis


4-3 Implementation Choices 43

Systematically decreasing Vref implies that the effective signal range in every step of the
binary search is diminishing. For example, if Vref drops to Vref ′ at the start of the 2nd bit
cycle in a binary search, the effective signal range available for the subsequent comparison is
′ /2 instead of V
Vref ref /2, as shown in Figure 4-11(a). This leads to incorrect digital outputs
in the same way as the incomplete reference settling does (Section 2-1-4). Such a systematic
decrease in Vref can be adjusted for by changing the capacitive division ratio. Instead of
performing a binary capacitive division by switching Cdac 2 at the start of the 2nd bit cycle,
Cdac
r is switched. Variable r is chosen so that the effective DAC voltage division seen by the
V
comparator is equal to ref ′
2 regardless of Vref dropping to Vref , as given by


Vref Vref
= . (4-6)
r 2
V′
As Figure 4-11(b) shows, by changing the capacitive division ratio to 2 Vref ref
, the required
signal range is covered. Similarly, the capacitive division ratio can be changed in every step
so that, regardless of the systematically decreasing Vref , the effective DAC reference levels
seen by the comparator follow a binary pattern (or a non-binary pattern as dictated by the
overrange required). However, it should be noted that changing r changes the charge pulled
′ . Choosing the right capacitive division ratio per step
from Cref , resulting in a different Vref
requires some design iterations.

Vref Vref

Vref Vref
2 2
Uncovered V′
ref
r=2 r = 2 Vref

0 0
i = MSB i = MSB - 1 i = MSB i = MSB - 1
time time
(a) (b)


Figure 4-11: Effect of Vref dropping to Vref on effective signal coverage (a) with conven-
tional capacitive division ratio, and (b) with appropriately changed capacitive division

4-3-2 Cancelling Charge

Removing only the code-dependent spread in Vref per bit cycle can be done in two ways as
shown in Figure 4-12. In the case of correcting ’upward’ as in Figure 4-12(a), the charge

Master of Science Thesis Iniyavan Elumalai


44 Charge-based Compensation of Reference Errors

required is supplied to Cref with a compensation capacitor Cc2,up charged to Vc , as given by



Cref Vref2,b + Cc2,up Vc = Cref + Cc2,up Vref2,a . (4-7)
Whereas, correcting ’downward’ (in Figure 4-12(b)) can be done by connecting an uncharged
capacitor Cc2,down such that it sinks the code-dependent part of the charge from the reference,
as given by  
Cref Vref2,a + Cc2,up .0 = Cref + Cc2,down Vref2,b . (4-8)

Vref

Vref2,a

Vref2,b

t1 t2 t3 t4 time
i=3 i=2 i=1 i=0

(a)

Vref

Vref2,a

Vref2,b

t1 t2 t3 t4 time
i=3 i=2 i=1 i=0

(b)

Figure 4-12: Removing the code-dependent spread in Vref (a) by adding charge, and (b)
by removing charge

Correcting ’downward’ essentially pulls more charge from reference in order to equalize
it over different codes. In order to enable this, the compensation bank has to be discharged
to ground initially instead of having to be pre-charged to Vc . This obviates the need for
providing Vc , which, being greater than Vref , would require a voltage booster in an actual
implementation. Interestingly enough, given this implementation, an analogy can be drawn
between the compensation bank and the capacitive DAC, in that switching a capacitor to the
reference from either of the arrays pulls code-dependent charge from the reference.

Iniyavan Elumalai Master of Science Thesis


4-4 Sources of Inaccuracy 45

4-3-3 Compensating for a Fewer Cycles

If the compensation technique discussed above is employed in each and every bit cycle, the
compensation bank and the digital logic performing selection may represent significant over-
head for high resolution systems. As evident by comparing the reference voltage spread after
compensation in Figures 4-3 and 4-4, compensating in an earlier bit cycle reduces the code-
dependent spread in the ensuing bit cycles by a considerable extent. Therefore, compensation
can be performed only for the first few bit cycles such that in the rest of the bit cycles, the
spread in reference without compensation is within the limits of reference error allowed, that
is, an LSB.
Summarizing the improvements to the compensation technique, Vref can be let to un-
dergo code-independent, systematic drops that are taken care of by changing the capacitive
division ratio per step. The code-dependent spreads in the first few bit cycles are removed
by appropriately sinking charge from Cref . The residual reference error in the rest of the bit
cycles can be tolerated by the ADC.

4-4 Sources of Inaccuracy

This section discusses the sources of inaccuracy in the proposed reference scheme. The salient
features of the proposed scheme, namely, changing the capacitive division ratio in order to
adjust for the systematically decreasing Vref , and charge-based compensation, are both based
on the fact that the charge pulled from the reference capacitor after every DAC switching
event is entirely deterministic. Any error source that challenges the accuracy in predicting
the reference charge lessens the effectiveness of this scheme.

4-4-1 Capacitor Mismatch

As the charge pulled from Cref is essentially determined by ratios of different capacitor values
in the DAC, the processing variations in these capacitor values directly affect the accuracy
of predicting the static errors in Vref . Figure 4-13 shows the results of 100 Monte Carlo
simulation runs in MATLAB, where the effect of DAC capacitor mismatch on the reference
voltage (as compared to Figure 4-1 without mismatch) for a 1 V, 8-bit ADC with Cref = Cdac
is shown. The unit DAC capacitance is assumed to be Cu = 1 fF, and each DAC capacitor is
modelled as
Ci = 2i−1 Cu + δi , (4-9)
where δi is a random variable with a zero mean and a variance of σi2 . The variance is given
by
σi2 = 2i−1 σu2 , (4-10)
where σu is the standard deviation of the unit capacitance [30], [31]. A 3% standard deviation
(i.e., (σu /Cu ) = 0.03) is assumed, which is a rather pessimistic assumption for a metal-finger
1 fF capacitor in 28-nm CMOS. As the figure shows, the deviation of the actual values for the
changes in Vref from the ones that would be predicted without mismatch varies considerably
with respect to DAC capacitor mismatches.

Master of Science Thesis Iniyavan Elumalai


46 Charge-based Compensation of Reference Errors

0.95

0.9

0.85

0.8
Vref

0.75

0.7

0.65

0.6

0.55
7 6 5 4 3 2 1 0
i (Bit cycle index)

Figure 4-13: Effect of DAC capacitor mismatch on the code-dependent errors in Vref
(simulated in MATLAB)

Furthermore, in the bit cycles where the code-dependent spread is removed by sinking
charge, a previously discharged capacitor that is valued appropriately for the code-dependent
spread in reference charge is connected to Cref . Besides the deviation in the reference charge
from the model due to DAC capacitor mismatch, random spread on the value of this com-
pensation capacitor implies that the compensation performed will not be accurate. It is to
be noted that an error in compensation in a bit cycle invalidates the compensation in the
next cycle, and so on. This is because, the charge pulled from Cref depends on the reference
voltage level before the DAC switching event. An inaccurate compensation in a bit cycle
results in a reference voltage level different from the desired value. This, in turn, causes the
reference charge in the next bit cycle to deviate from the model.
Processing variations in the DAC and compensation capacitors can greatly affect the
accuracy of the proposed reference scheme. In order to be able to predict the errors on
the reference with 14-bit accuracy, the DAC capacitors should match to at least 0.0064%.
At present, accuracies up to 12 bits can be expected (in the cases of most of the switch-
ing algorithms [31]) by employing high-quality metal-insulator-metal (MIM) or metal-finger
capacitors with careful layout measures [12]. Further accuracy is by virtue of foreground
and/or background digital calibration algorithms that correct for DAC capacitor mismatches
[21, 19, 32, 33, 34, 35]. Correction parameters determined by such algorithms can be used to
calibrate the errors in the predicted reference model.

4-4-2 Switch Nonidealities

Switches that are tied to Cref , namely, the switches connecting to the DAC capacitors and
those to the compensation capacitors introduce errors in Vref through two distinct mecha-
nisms. Firstly, these switches couple their clock transitions to Cref through their overlap

Iniyavan Elumalai Master of Science Thesis


4-5 Reference Scheme Implementation in the Fine SAR ADC 47

capacitances as they turn on and off. Secondly, these switches inject their channel charge into
Cref as they turn off, which appear as negative or positive pedestals on Vref (with NMOS
and PMOS switches respectively). These errors in Vref , although code-independent, affect
the accuracy of the reference scheme in that they affect the predictability of reference charge.
Sizing the MOS switches with small aspect ratios helps reduce the errors due to switch
nonidealities. However, smaller switches have high on-resistance which reduces the DAC
settling speed. Techniques using dummy or complementary switches are usually employed
in order to circumvent this speed-precision trade-off. Using dummy switches to absorb the
charge injected is less attractive, as the dummies need to be sized carefully in order to ensure
exact removal of the fraction of channel charge injected into Cref , which is generally
cumbersome [36]. The approach incorporating complementary switches, such that the op-
posite charge packets injected by the two cancel each other, requires the complementary
switches (NMOS and PMOS) that are matched to each other. Furthermore, this approach
is only good for one input level [13]. Since Vref is systematically decreasing in the proposed
reference scheme, the channel charge injected varies every bit cycle, making this approach
unsuitable.
A simple, viable solution is to include the excess charge introduced by the switches
into the model of the reference charge. Designing the capacitive radix, and choosing the
compensation capacitors can then be performed accordingly. However, processing variations
in MOS switches that introduce randomness in the excess charge still affect the accuracy of
the reference scheme.
Besides the aforementioned sources, power supply variations and capacitor parasitics also
affect the accuracy in predicting the reference charge. Enough power supply rejection should
be provided while charging the reference capacitor during the Sample phase. For the latter,
the compensation capacitors can be designed taking into account the simulator-estimated
parasitics.

4-5 Reference Scheme Implementation in the Fine SAR ADC

The reference scheme discussed so far, being bound in accuracy by the processing variations
of different capacitors and switches present in the ADC, becomes limited to use in medium-
resolution converters. In the prototype two-stage, SAR-assisted pipeline ADC, the fine SAR
ADC can afford relaxed precision because of the interstage gain. The precision requirements
in the coarse SAR ADC are the same as those of the entire ADC. Therefore, the proposed
reference scheme is implemented in the fine SAR ADC. This section discusses the circuit-level
implementation details and the results of the reference correction technique investigated.
The fine ADC, with a rail-to-rail input range of 375 mVpp,dif f , has to resolve 8 bits
effectively. Vref p and Vref n are chosen to be 0.8 V and 0.5 V, respectively. They are attenuated
by connecting a fraction (= 3/8) of the total sampling capacitance, Cs to ground during the
Convert phase, resulting in an effective Vref = 187.5 mV that covers the entire signal range.
The capacitive DAC employs the set-and-down switching algorithm. During the Sample
phase, the input is sampled on the top plates of the array capacitors (tied to the input of
the comparator), while their bottom plates are connected to VCM = 0.5 V. At the start of
the Convert phase, the following actions are performed: 1) the top plates are disconnected

Master of Science Thesis Iniyavan Elumalai


48 Charge-based Compensation of Reference Errors

from the input, 2) 38 Cs is connected to the ground, and 3) the bottom plates of the remaining
capacitors in the array (whose total capacitance is equal to Cdac ) are all connected to Vref p .
At the end of this first bit cycle the result of the MSB comparison is available. As the next
bit cycle begins, the bottom plate of the MSB capacitor in either the positive or the negative
DAC (based on the MSB decision) is switched from Vref p to Vref n . This results in an effective
DAC step of Vref /x, where x is the radix of the search through the DAC. Further DAC
switching proceeds according to the algorithm until 8 bits are resolved (Section 3-3).

VCM
Vref p
Vref n

CN −1 CN −2 C0
Vdacp
Vip

Figure 4-14: Single-ended schematic of the capacitive DAC in the fine stage

Vdacp ,
Vdacn
Vref Vref
(Vref − VCM ) r r2
Vref
Vip r3

Vin

Phase: Sample Convert time

Figure 4-15: Examples of positive and negative DAC voltage waveforms in the fine stage

Figure 4-14 shows a simplified, single-ended version of the DAC with the different switches
involved. Figure 4-15 shows typical waveforms of the positive and negative voltages input
to the comparator. It can be noticed that the jump from VCM to Vref p in the first step is
performed in order to shift the common-mode level of the voltage present at the comparator
input from that of the sampled input (dashed grey lines represent the CM levels). This step
pulls charge from the reference capacitor, which had not been modelled earlier in Figure 4-1.

Iniyavan Elumalai Master of Science Thesis


4-5 Reference Scheme Implementation in the Fine SAR ADC 49

Since the consequent error on Vref is signal-independent, it is considered to be an offset, that


is taken care of by changing the reference attenuation factor.
The proposed reference scheme is implemented with Cref p = Cref n = 1 pF. The capacitive
DAC is designed such that the search radix takes into account the systematically decreasing
Vref . Overrange is also built in to take care of the errors due to processing variations and
residual code-dependence in the bit cycles without compensation. After a number of design
iterations, the final DAC is designed to consist of 10 capacitors whose weights are given by
{66 36 19 10 7 4 3 2 1 1}. With a unit capacitance (Cu ) chosen as 0.7 fF, the total DAC
capacitance is equal to 100 fF. Figure 4-16 shows the DAC signal range considered during
every step of the search. It should be noted that the figure takes into account the code-
independent drops in Vref and the nominal switch charge injection in every bit cycle. With
an average radix of 1.6, it takes 11 bit cycles, including the step involving CM shift, to reach
an ENOB of 8 bits.
Charge-based compensation is performed in the first three bit cycles with code-dependent
spreads (i = 7, 6 and 5) with a bank of compensation capacitors that amount to a total
capacitance of 20 fF. The residual code-dependent spread in Vref at the end of the last bit
cycle is 0.62 mV, which is less than one LSB (= 1.465 mV). The worst-case spread over
process/temperature variations and over 7 different Monte Carlo runs are also less than one
LSB.

1.8
Effective signal range per bit cycle

1.6
normalized to V = 187.5 mV

1.4

1.2
ref

0.8

0.6

0.4

0.2

0
9 8 7 6 5 4 3 2 1 0
i (Bit cycle index)

Figure 4-16: Effective signal coverage per bit cycle with the designed cap. radix, taking
into account the code-independent drops in Vref and nominal switch charge injection in every
bit cycle

Master of Science Thesis Iniyavan Elumalai


50 Charge-based Compensation of Reference Errors

4-6 Simulation Results

Figure 4-17 shows the 4096-point FFT output for the fine SAR ADC with the designed
reference scheme sampling a 90% FS, 3.1 MHz input signal at 100 MS/s. Total SNDR
including all the noise and distortion components within the Nyquist bandwidth is 50.9 dB,
which corresponds to an ENOB of 8.1 bits. Values of SFDR and THD in a typical corner are
C
61.8 dB and 60.5 dB respectively. Compared to Figure 4-4 for the same Crefdac
ratio (= 10), the
implemented reference correction technique thus reduced the reference-induced distortion by
19.5 dB in SFDR and 21.6 dB in THD. It should be noted that the reference-induced distortion
is reduced to a little more than what is necessary for 8-bit linearity. Even though further
improvement can be expected by not making the simplified implementation choices discussed
in Section 4-3, it is rather important to mention that the effectiveness of this technique is
primarily limited by the processing variations in capacitors and MOS switches.

−10
SNDR = 50.9 dB
−20 ENOB = 8.1 bits

−30
Amplitude [dB]

−40
Quantization noise floor
−50

−60
FFT Gain = 10 log(4096/2)
−70

−80 FFT noise floor

−90

−100
0 10 20 30 40 50
Frequency [MHz]

Figure 4-17: Fine ADC performance with the designed reference scheme including charge-
based compensation in 3 bit cycles, and changed capacitive radix to accommodate the code-
independent drop in Vref

4-7 Summary

In this chapter, we studied the ADC distortion caused by code-dependent errors in Vref .
A charge-based compensation technique taking advantage of the predictability of reference
errors was proposed to deal with this reference-induced distortion. Finally, we looked at the
proposed reference scheme being implemented in the fine stage of the prototype two-stage,
pipelined SAR ADC.

Iniyavan Elumalai Master of Science Thesis


Chapter 5

Digital Correction of Reference


Errors

In this chapter, we propose a digital calibration technique to correct for nonlinear errors
arising from the code-dependence in Vref held by a stand-alone capacitor. The implementation
details of the proposed technique in the coarse stage of the prototype, two-stage, SAR-assisted
pipeline ADC are presented in detail. The design methodologies of the critical building blocks
in the proposed reference scheme are discussed, followed by simulation results.

5-1 Need for Digital Calibration


The charge-based compensation of reference errors discussed in the previous chapter, being an
error-correction technique in the analog domain, becomes ineffective at resolutions of 12 bits
or higher thanks to inherent analog impairments, namely, processing variations in capacitors
and switches. Moreover, this technique involves a bank of compensation capacitors which,
when sized for accuracy requirements and designed for compensation in every bit cycle, can
occupy chip area as large as a significant fraction of the area occupied by the main DAC. The
logic performing the selection from the bank alongside corresponding switches also represents
a major overhead. This necessitates a correction technique that can help remove the reference-
induced distortion up to 14-bit precision or higher, meanwhile representing minimal overhead.
In that regard, this chapter proposes a new approach that seeks digital assistance to deal
with the code-dependent errors in Vref held by a stand-alone capacitor. This approach of
resorting to digital calibration techniques in order to improve the ADC linearity is prominent
in the recent literature. Various digital calibration algorithms have been proposed for SAR-
based converters to correct for nonlinear factors such as capacitor mismatch [37, 19, 32,
33, 34, 35], and in the case of a pipelined architecture, inter-stage amplifier nonlinearity
[38, 39, 40, 41, 42, 43]. This growing prevalence of digital assistance is fuelled by the fact
that the aggressive realization of Moore’s law over the past few years has greatly benefited
digital circuit performance (eg., in speed and power), resulting in fast, efficient digital cores
at smaller costs.

Master of Science Thesis 51 Iniyavan Elumalai


52 Digital Correction of Reference Errors

5-2 Calibration Methodology

The digital calibration algorithm proposed here derives its inspiration from a popular calibra-
tion technique used in pipelined ADCs that extracts non-idealities such as capacitor mismatch,
finite op-amp gain, comparator mismatch, etc., by processing the conversion residues from the
coarse stages. The idea that the residue voltage contains information about the non-idealities
of the conversion stage, and that it can be processed in a pipeline architecture to extract and
correct such imperfections was first explored in [44]. The algorithm reported here claims that
the reference errors in a coarse stage appear in the residue voltage in the same way as any of
the aforementioned non-idealities would. Therefore, taking advantage of the pipeline archi-
tecture of the prototype ADC, this algorithm is employed to correct for the reference errors in
the coarse stage, which simply uses a medium-valued stand-alone reference capacitor with no
correction in analog domain, by processing its residue. The goal is to achieve at least 14-bit
linearity with respect to reference-induced distortion, while guaranteeing minimal overhead.

5-2-1 Residue Characteristic

In order to develop the algorithm for the correction of reference errors, their effect on the
residue characteristic is first formulated. Figure 5-1 shows the DAC tree for a 2-bit, differ-
ential, binary search with all the ideal reference levels per bit cycle. The full-scale range
2V
{−Vref , Vref } is divided into 22 bins, each bin the size of an LSB (= 2ref2 ). For an input
Vi located in the bin j, an ideal binary search follows the corresponding trajectory (which
represents the digital output code) as shown in the figure. The voltage residual after the
conversion is given by
Vres = Vi − Vdacend , (5-1)
where Vdacend is the final reference level created at the end of the conversion cycle after
switching the LSB capacitor. It is given by

2j − 22 + 1
Vdacend = Vref (5-2)
22
for j = 0, 1, ..., 22 − 1,

with j representing the height of the digital output code. The residue voltage, when plotted
against the input voltage, gives the ideal residue characteristic as shown in Figure 5-2. The
sawtooth in each bin has a constant height and base, both equal to 1 LSB.
After including the effect of reference errors in the case of a stand-alone reference ca-
pacitor, the reference levels become erroneous as shown in Figure 5-3. The residue voltage
becomes
2j − 22 + 1
Vres = Vin − Vrefend,α (5-3)
23
for j = 0, 1, ..., 22 − 1,

where Vrefend,α represents the reference voltage on Cref after the LSB capacitor is switched.
Parameter α implies that this voltage varies in a quadratic fashion with digital output code,
that is, according to j. The resulting residue characteristic is as shown in Figure 5-4.

Iniyavan Elumalai Master of Science Thesis


5-2 Calibration Methodology 53

+Vref

j=3

j=2

Vdac 0
j=1

j=0 LSB =
2Vref
22
−Vref

D1 D0 time

Figure 5-1: Ideal DAC tree for a 2-bit, differential search

Vres

(Vref , Vref )

Vin
(−Vref , −Vref )

LSB

LSB

Figure 5-2: Ideal residue characteristic

Master of Science Thesis Iniyavan Elumalai


54 Digital Correction of Reference Errors

+Vref

Vdac 0

−Vref

D1 D0 time

Figure 5-3: Effect of reference errors on the DAC tree

Vres

(Vref , Vref )

Vin
(−Vref , −Vref )

X 6= Y 6= LSB

Figure 5-4: Effect of reference errors on the residue characteristic

It becomes apparent from the figures that the LSB spacings on the DAC tree, affected
by reference errors, become uneven. To that end, the calibration algorithm reported here
attempts to determine the relative LSB spacings on the DAC tree, and therefore estimate the
erroneous DAC transfer, from which the errors on the reference can be corrected.

Iniyavan Elumalai Master of Science Thesis


5-2 Calibration Methodology 55

5-2-2 Calibration Algorithm

The proposed algorithm is similar to [38] in that an additional step is introduced after the
conversion where the DAC voltage can be modulated to result in two different residue modes.
In [38], two different residue modes are randomly alternated, and residual differences are
obtained that serve to estimate and correct for the inter-stage gain errors. In the algorithm
used in this work, such residual differences obtained for different trajectories are used to
correct for the reference-induced errors in the DAC transfer.

Vi − Vdac
Three possible residues

0 LSB
time
D1 D0

Extra step

Figure 5-5: DAC voltage modulation in an additional step after SAR conversion

Figure 5-5 shows a typical waveform of the voltage present at the comparator input. After
the conversion has ended, that is, the LSB capacitor has been switched, the DAC voltage is
modulated. The modulation is performed by including an additional capacitor in the DAC
array. Switching this capacitor gives the same step-size as that of the LSB step. As a result,
there are three different possible residue voltages corresponding to the additional capacitor
taking no step, a positive step (Vref p to Vref n ) or a negative step (Vref n to Vref p ), as shown
in the figure.

Figure 5-6 shows the DAC tree with three different cases of modulation (the DAC voltages
in the case of no modulation are not shown in the additional step for the sake of clarity). In
order to estimate the relative LSB spacings on the DAC tree (as shown in the figure) values
of h1 (j) and h2 (j) should be computed for every trajectory j throughout the full-scale range,
as follows:

1. h1 (j) represents the size of the output LSB bin corresponding to the trajectory j. As
Figure 5-7 illustrates, h1 (j) can be obtained by determining the difference between the residue
voltages corresponding to positive and negative DAC voltage modulations for the trajectory
j at a given input voltage V1 (j) in bin j.

Master of Science Thesis Iniyavan Elumalai


56 Digital Correction of Reference Errors

+Vref

h1 (j)

h2 (j)

Vdac 0

−Vref

D1 D0 time

Figure 5-6: DAC tree with three different cases of DAC voltage modulation

Vres
j j+1

h1 (j)
Vin

V1 (j)
Figure 5-7: Determining h1 (j)

2. h2 (j) represents the spacing between the LSB bins j and j+1 on the DAC tree, and
obtaining this through out the full-scale gives the relative locations of the LSB bins.
Figure 5-8 shows the residue voltage in the case of no DAC voltage modulation in the ad-
ditional step. It can be observed from the figure that h2 (j) is the difference between the
two residue values that are obtained by fixing the input voltage at V2 (j) and forcing the
trajectories j and j+1 in the SAR loop.

Iniyavan Elumalai Master of Science Thesis


5-2 Calibration Methodology 57

Vres

j j+1
h2 (j)
Vin

V2 (j)

Figure 5-8: Determining h2 (j)

These differences are formed in the digital domain after the residue voltages concerned
have been digitised by the back-end. Once h1 (j) and h2 (j) have been obtained for each and
every trajectory, the complete DAC transfer can be constructed by stacking the estimated
LSB bins on top of each other. Following this, INL-based correction terms are obtained by
finding the deviation of the estimated transfer from an ideal ’best-fit’ transfer.

5-2-3 Deterministic Calibration Procedure

This section analyses different possible ways of implementing a digital calibration algorithm,
and derives a convenient, deterministic procedure for the algorithm discussed above. In gen-
eral, operability in the background and fast computation of the required correction parameters
are the two desirable features for any ADC calibration scheme. Without interrupting the nor-
mal ADC operation, the calibration procedure should be short enough to track parameter
variations due to environmental influences such as temperature. However, a review of the re-
lated literature shows that these features are somewhat mutually exclusive, making it difficult
to develop an ideal calibration procedure with minimal overhead.
A complete background calibration procedure that does not impose any restrictions on
the input signal characteristics usually involves randomization of the non-linear error terms
with a pseudo-random noise (PN) sequence, as evident in many of the previously reported
ADC calibration schemes [38, 45, 46, 34]. This idea can be extended to the foregoing al-
gorithm, where the DAC voltage in the additional step can be modulated based on a PN
sequence. This way, the reference errors will become PN-modulated, and multiplying the
digital output by the same PN sequence will help decorrelate and extract the information
required to correct reference-induced nonlinearity. However, such a scheme will suffer from a
major trade-off between measurement accuracy and the computation time. This is because
in order to extract the required information from the ADC output with high precision, a
large number of conversions have to be performed to average out the input. Since the PN-
modulated input content present in the √ digital output can be considered as a white noise
whose standard deviation reduces by 2 as the the number of averaging samples is doubled

Master of Science Thesis Iniyavan Elumalai


58 Digital Correction of Reference Errors

[47], 22N conversions are required to achieve N-bit accuracy. In order to determine the refer-
ence errors in the coarse stage of the prototype ADC resolving 7 bits, 128 trajectories should
be tested. For the whole calibration to be 14-bit accurate, each test should be at least 17.5-bit
accurate [48]. Thus, a background calibration scheme would require 128 × 22×17.5 samples, or
∼12 hours of conversion time with a 100 MS/s sampling clock.
The split-ADC approach [49], a parallel, slow-but-accurate ADC [33], and signal-dependent
dithering [47] are some of the techniques from recent calibration literature that attempt to
achieve higher accuracy and faster calibration cycles simultaneously. However, they mostly
incur additional analog and/or digital complexity [43]. Rapid conversion cycles are possible
with minimal overhead in a simple, ’deterministic’ calibration scheme, as evident in the cali-
bration schemes from the early 90s [44, 50, 51]. The limitation with a deterministic scheme
is that since it is based on providing a known, analog test signal to the ADC, the ADC is
required to go off-line during the calibration procedure, which inhibits its use in the complete
background. However, in ADC architectures with possibilities for some forms of redundancy,
such as additional ADC lanes or extra conversion time, a deterministic calibration scheme
can be employed without such a limitation. The prototype ADC is implemented with a
time-interleaved architecture, thereby allowing an additional lane to accommodate a deter-
ministic calibration scheme. It should be noted that measuring h2 (j) is possible only with a
deterministic scheme.
In order to enable a deterministic scheme for the above-said calibration algorithm, the
test signals, namely, V1 (j) and V2 (j) should be generated in order to help obtain the residual
differences h1 (j) and h2 (j) for every trajectory j. This can be accomplished with the help of a
calibration DAC, the linearity and speed requirements of which are discussed later in Section
5-3-2.

5-3 Reference Scheme Implementation in Coarse SAR ADC

This section presents the implementation details of the reference scheme in the coarse ADC
of the prototype.

5-3-1 System-level Details

The prototype ADC consists of two ADC lanes, one of which is used once a second for reference
calibration. Even though this seems to imply huge penalties in terms of area and power, it
should be mentioned that this research mainly attempts to prove the concept of reference
calibration. In an actual industry-oriented chip with a time-interleaving factor greater than 2,
the penalties of taking a single lane off-line can be considered negligible. The timing diagram
is shown in Figure 5-9. The main conversion lane operates with the sampling clock of 100
MHz and 50% duty-cycle between Sample (φ1 ) and Convert (φ2 ) phases. The calibration
lane is used once a second to perform the conversions required for the complete calibration
procedure. During that period, the calibration lane samples its test input signals from a
calibration DAC which has an update rate of 10 MHz. Thus, the clock for the calibration
lane is 5% duty-cycled as shown in the figure.

Iniyavan Elumalai Master of Science Thesis


5-3 Reference Scheme Implementation in Coarse SAR ADC 59

10 ns

SAR φ1 φ2
100 ns

Calibration φ1 φ2

Figure 5-9: Timing for the two ADC lanes in the prototype during calibration

Vcalp SAR engine


Vip
Vref p mode
Vref n
FF
Q R cmp_out
hii
CN −1,1 CN −1,2 CN −2,1 CN −2,2 C0,1 C0,2 S
...
selhii

Vdacp

VCM

Figure 5-10: Simplified block diagram of the coarse ADC

Figure 5-10 shows a simplified block diagram of the coarse ADC with its essential building
blocks, namely, a capacitive DAC, a comparator and the SAR logic loop alongside calibration
control. There are two modes of operation:

1. During the SAR mode, the ADC operates in a normal fashion with the capacitive
DAC sampling the input and the SAR logic proceeding with corresponding comparator out-
puts.

2. During the Calibration mode, the ADC samples test input voltages from the calibra-
tion DAC in order to obtain the difference estimates h1 and h2 for each and every trajectory.
For the sake of flexibility, each trajectory can actually be forced in the SAR loop by taking the
comparator off-line, and feeding the SAR engine with inputs from latches (FF, as shown in
the figure) that are controlled by MATLAB. This flexibility is indispensable because depend-
ing on the linearity of the calibration DAC, the test voltage corresponding to the trajectory
under test may be off the desired bin, making the residue voltage exceed the range allowed
by the back-end. By observing the output from the fine ADC with external control, tests can
be modified for each trajectory until all the information required to estimate the entire DAC
transfer are obtained. It should be remarked here that an easily-integrable implementation
of this calibration control has yet to be designed.

Master of Science Thesis Iniyavan Elumalai


60 Digital Correction of Reference Errors

SSAR
RSAR Q

Scal
Rcal

Figure 5-11: Schematic of the latch in the main SAR loop

Table 5-1: SR inputs during different modes of operation

Mode SSAR Scal RSAR Rcal


SAR clk_conv 1 cmp_out↑selhii 1
Calibration 1 clk_conv 1 FF ↑selhii

The said multiplexing in the SAR engine is implemented with the help of a four-input
NAND-based SR latch, as shown in Figure 5-11. SSAR and RSAR are the active inputs during
the SAR mode, whereas Scal and Rcal are active during the Calibration mode. Their values
during the two modes are given in Table 5-1. clk_conv is used to preset the latch at the start
of every conversion cycle, and selhii is used to enable the latch i during the ith bit cycle.

5-3-2 Circuit-level Details

The coarse ADC resolves the input full-scale range of 1.4 Vppd into 7 bits. With the reference
voltages being Vref p of 1 V and Vref n of 0 V, reference attenuation is performed as explained
in Section 3-5 to obtain the effective Vref of 0.7 V.

Capacitive DAC

The capacitive DAC employs split-cap switching algorithm [24], where every capacitor in the
array Ci is split into two sub-capacitors Ci,1 and Ci,2 , as shown in Figure 5-10. During the
Sample phase, the top plates of all of the array capacitors tied to the input of the comparator
are connected to VCM , whereas the input is sampled on the bottom plates. At the start of
the Convert phase, the bottom plates of every Ci,1 are connected to Vref p , and those of Ci,2
are connected to Vref n . This sets up the DAC voltage for the MSB comparison. As the next
bit cycle begins, depending on the result of the MSB comparison, either CN −1,1 or CN −1,2
is toggled differentially. For a positive decision, CN −1,2 in the positive DAC is connected to
Vref p , and CN −1,1 in the negative DAC is connected to Vref n . For a negative decision, CN −1,1
in the positive DAC is connected to Vref n , and CN −1,2 in the negative DAC is connected to
Vref p . Further DAC switching proceeds in the same way. A clear illustration of the split-cap
switching algorithm is given in Appendix A.
The proposed reference scheme is implemented with a Cref of 20 pF. Even though the
aforementioned calibration technique is implemented to take care of the non-linearity due
to code-dependent errors on the reference voltage, the ADC can not handle significantly
decreasing Vref because the residue voltage at the end of the coarse conversion may exceed

Iniyavan Elumalai Master of Science Thesis


5-3 Reference Scheme Implementation in Coarse SAR ADC 61

the range that can be handled by the back-end (the residue amplifier and the fine SAR).
Therefore, the capacitive DAC is designed such that the capacitive division per bit cycle
accounts for the code-independent drops in Vref , as discussed in Section 4-3-1. The weights
of the sub-capacitors in a single-ended array are given by {34, 19, 11, 7, 4, 3, 2, 1, 1{. With
unit capacitor chosen as 7 fF, the total DAC capacitance equals 1.1 pF. With an average
radix of 1.52, it takes 10 bit cycles (with the MSB step resolving 1 bit) to reach an effective
resolution of 7 bits.

Pre-charge Buffer

With a 20 pF reference capacitor and the designed capacitive DAC, the code-dependent spread
on Vref at the end of SAR conversion is determined to be 10 mV. This residual spread must
be removed before the start of the next conversion up until 14-bit precision, that is, with
an allowed error of 85.5 µV (= 21.4
14 ). For that purpose, a pre-charge buffer replenishes Cref
during the Sample phase. With a 100 MHz sampling clock, 14-bit precise settling has to be
done in 5 ns, as given by  
10 mV
5 ns = Rout Cref ln s. (5-4)
85.5 uV
This requires the output impedance of the buffer to be less than 50 Ω.

M2

out Zout 1
∼ gm
in M1
1
2 r
gm o

ω
1 gm2
ro1 Cgs2 Cgs2

(a) Flipped voltage follower (b) Output impedance versus frequency

Figure 5-12: Flipped voltage follower and its output impedance versus frequency

Such a low output impedance can be achieved with a source follower (Rout = 1/gm ),
which, however, would require a large class-A current. Negative feedback can be employed to
reduce the output impedance further. Figure 5-12(a) shows the circuit diagram of a flipped
voltage follower, which can be described as a voltage follower with shunt feedback [52]. Its
output impedance is derived as follows. When the output node suffers a voltage change of
∆V , M1 acts as a common-gate stage leading to an amplified voltage change of gm1 ro1 ∆V at
the gate of M2. M2, consequently, enables a change of gm2 (gm1 ro1 ∆V ) in the output current,
resulting in an output impedance given by
1
Rout = . (5-5)
gm1 ro1 gm2

Master of Science Thesis Iniyavan Elumalai


62 Digital Correction of Reference Errors

Thus, with this voltage-sensed, current feedback, the output impedance of a simple source
follower is reduced by an amount equal to the intrinsic gain (= gm ro ) at low frequencies.
Figure 5-12(b) shows the output impedance versus frequency. It should be remarked here
that this circuit exhibits excellent current sourcing capability, as the required load current is
provided from the supply through M2, while M1 conducts constant current. This suits the
requirement for the pre-charge buffer, as the buffer is expected to supply current in only one
direction. In order to pre-discharge a capacitor holding the negative reference voltage, an
NMOS version of this circuit can be employed for its excellent current sinking capability.
The implemented pre-charge buffer is shown in Figure 5-13. M1 and M2 constitute the
enhanced source follower. M4 cascodes M3 to form the tail current source. M5 and M6 shield
the output node from power supply variations, together providing 40 dB of rejection. It can
be noticed that for Vref to be 1 V from a 1.8 V supply, the headroom is limited for all the
transistors to operate in saturation. A parallel RC level shifter (R1 ||C1 ) is placed inside the
feedback loop which offers the headroom required [53]. S1 and S2 are the switches connecting
the buffer to Cref . They are placed inside the loop so that they do not degrade the effective
output impedance seen by Cref .
A slow loop is employed to improve the DC accuracy of Vref . It consists of an error
amplifier that adjusts the input to M1 after comparing Vref with the bandgap reference. In
order to decouple the slow loop from the output node glitches, the slow loop senses the output
voltage only for the second half of the Sample phase, after the transients have settled to a
significant precision. The implemented error amplifier is a single, differential stage with a
PMOS-input topology. With a bias current of 0.3 µA, it achieves a DC gain of 50 dB and
a unity-gain bandwidth of 65 kHz. A large capacitance (= 10 pF) is used at its output to
suppress its noise contribution.
Since the buffer is not being used during the Convert phase, power-cycling is employed.
It is advantageous that this buffer exhibits fast turn-on transients. The buffer consumes a
bias current of 50 µA during the Sample phase, and 3 µA during the Convert phase. Figure
5-14 shows the transient waveforms of Vref in blue corresponding to two digital codes that
have the maximum and minimum errors in Vref at the end of SAR conversion. It can be seen
that the code-dependent spread is removed up to 14-bit precision during the Convert phase
of 5 ns. The figure also includes waveforms of currents from the supply for these two extreme
cases. The buffer responds to the voltage steps with currents as large as 1 mA. The average
current consumed by the pre-charge buffer, including the biasing blocks, is 165 µA from a 1.8
V supply.

Calibration DAC

The calibration DAC generates the test voltages required (V1 (j) and V2 (j)) during the cali-
bration procedure. To derive the linearity requirements of the DAC, the allowable deviations
of V1 (j) and V2 (j) from ideal values should be known. While testing for the trajectory j,
as Figure 5-15 shows, V1 (j) should be located within the input bin j, and V2 (j), within the
hatched region around the trip point between bins j and j + 1. Otherwise, the residue volt-
ages exceed the conversion range of the back-end. Since the allowable deviations of the test
voltages are up to the size of an LSB at the 7-bit level, the linearity requirements of the DAC
can be relaxed. Moreover, for the sake of flexibility during the calibration procedure, the test

Iniyavan Elumalai Master of Science Thesis


5-3 Reference Scheme Implementation in Coarse SAR ADC 63

Vb6
M6

Vb5
M5
Ib

M2

S1
R1 C1 S2 Vref
S1
S3 Cref
Vref
M1
Bandgap
Cslow

Ib Vb4
M4

S1
Vb3,a
Vb3,b M3
S2

(a) Implemented buffer schematic

Sample Convert Sample

S1

S2

S3

(b) Timing for the switches involved

Figure 5-13: Schematic of the implemented pre-charge buffer

voltages are generated at 9-bit resolution, offering many test voltages per trajectory. The
specifications for the calibration DAC are given in Table 5-2.
A 9-bit, current-steering DAC is implemented. The architecture is segmented "6 + 3",
where the 6-bit sub-DAC is for the MSBs and the other for the LSBs. While the LSB section
is binary-coded, thermometer coding is used in the MSB section which helps achieve good
DNL performance [54]. Since the DAC is used only for calibration purposes, not much design
effort was invested in optimizing power and high-frequency linearity performance.

Master of Science Thesis Iniyavan Elumalai


64 Digital Correction of Reference Errors

Iref [mA]
Vref [V]

0.5

0.95
50 µA 3 µA
10 mV

Precharge phase (5 ns) 0


4.01 4.011 4.012 4.013 4.014 4.015
time [s] x 10
−6

Figure 5-14: Transient response of the implemented pre-charge buffer

Vres Vres
j j+1 j j+1

Vin Vin

V1 (j) V2 (j)

(a) Allowed location for V1 (j) (b) Allowed location for V2 (j)

Figure 5-15: Allowed locations for the test voltages required during calibration

With the full-scale current equal to 225 uA, the unit LSB current is given by 440 nA. The
designed unit current cell is shown in Figure 5-16. It consists of a current source transistor
Mcs , a cascode transistor Mcas and a pair of differential switches Msw . Since the matching of
current sources is important for DNL performance, a large current source transistor is used
[55]. With feature sizes at Wcs = 1.4µm and Lcs = 0.75µm, the standard deviation on a unit
current source over 100 Monte Carlo runs is determined to be 0.08 LSB.
Since this DAC operates from a 1.8 V power supply, the amount of voltage headroom
left above the current steering circuitry is not sufficient to provide the required output swing

Iniyavan Elumalai Master of Science Thesis


5-3 Reference Scheme Implementation in Coarse SAR ADC 65

Table 5-2: Specifications for the calibration DAC

Resolution 9 bits
Update Frequency 10 MHz
|DNL| < 0.5 LSB
|INL| < set by DNL
Output Swing 1.5 Vppd
Output CM-level 0.5 V

z̄ Msw z

Vcas Mcas

Vcs Mcs

Figure 5-16: Schematic of a unit current cell

1.8 V

2:1 1:2

Ibleed Ioutp Ioutn Ibleed


Vop Von
Vcm
R1 R2

To ADC To ADC

Figure 5-17: Schematic of the output network in the calibration DAC

of 1.5 Vppd. This headroom limitation is circumvented by mirroring the output currents from
the current cells, as shown in Figure 5-17. The cascode current mirror is designed to produce

Master of Science Thesis Iniyavan Elumalai


66 Digital Correction of Reference Errors

linear output currents while ensuring sufficient headroom [48].

0.8

0.6
Differential output voltage [V]

0.4

0.2

−0.2

−0.4

−0.6

−0.8
0 100 200 300 400 500
Input code

Figure 5-18: Transfer function of the implemented calibration DAC

LSBs at 9−bit level


1

0.5
INL (LSB)

−0.5

−1
0 100 200 300 400 500

0.5
DNL (LSB)

−0.5

−1
0 100 200 300 400 500
Input code

Figure 5-19: Worst-case INL/ DNL performance of the implemented calibration DAC

The common-mode (CM) level at the output is defined and maintained by the DC bleed

Iniyavan Elumalai Master of Science Thesis


5-4 Simulation Results 67

current Ibleed and the common-mode feedback network shown in red. Differential resistors R1
and R2 sense the CM level of the two outputs. The measured CM level is compared with the
reference using an error amplifier, and the error is returned to the output bias leg. The error
amplifier implemented is a single, differential stage, with a DC gain of 20 dB.
The simulated DAC transfer curve is shown in Figure 5-18. Figure 5-19 shows the worst-
case INL and DNL measured over 5 different Monte Carlo simulations. The DNL plot shows
that the DAC is 9-bit monotonic, which is more than sufficient for the calibration procedure.

5-4 Simulation Results

This section gives the simulation results for the reference scheme implemented in the coarse
SAR. As mentioned earlier, the reference-induced distortion shows up as third-order non-
linearity in the capacitive DAC transfer, and thus the whole ADC transfer. This is evident
in Figure 5-20, where the INL of the capacitive DAC is plotted in LSB values at the 14-bit
level versus the 7-bit output code of the coarse SAR. Figure 5-21 shows the 128-point FFT
output for the coarse ADC with the designed reference scheme sampling a 90% FS, 2.43475
MHz input signal at 100 MS/s. It should be noted here that this spectrum is obtained of the
7-bit coarse code along with the conversion residue. The reference-induced harmonic is 55 dB
below the fundamental, limiting the linearity to around the 9-bit level.

LSBs at 14−bit level


15

10

5
INL (LSB)

−5

−10

−15
0 20 40 60 80 100 120
Output code

Figure 5-20: INL of the capacitive DAC with Cref = 20Cdac before calibration

Since the capacitive DAC has overrange built in, there are multiple possible trajectories
that represent the same digital output which complicates the calibration procedure. However,
it can be assumed that such equivalent trajectories have the same residual error in Vref
since they travel the same distance along the DAC tree. Thus, the calibration procedure is
performed only for the 128 trajectories that represent 128 different digital outputs. Figure
5-22 shows the INL of the capacitive DAC after calibration. The reason for the residual error

Master of Science Thesis Iniyavan Elumalai


68 Digital Correction of Reference Errors

Quantization residue included


0

SNDR = 54.61 dB
−20
SFDR = 54.64 dB

−40
Amplitude [dB]

−60

−80

−100

−120
0 10 20 30 40 50
Frequency [MHz]

Figure 5-21: Coarse ADC performance with Cref = 20Cdac before calibration

LSBs at 14−bit level


1

0.8

0.6

0.4

0.2
INL (LSB)

−0.2

−0.4

−0.6

−0.8

−1
0 20 40 60 80 100 120
Output code

Figure 5-22: INL of the capacitive DAC with Cref = 20Cdac after calibration

is that this calibration algorithm employs the distance estimates that are obtained in the
additional step to correct for the reference errors during SAR conversion. The error caused
in Vref during this additional switching is unaccounted for. The sine-wave performance after
calibration is shown in Figure 5-23. Thus, the reported calibration scheme brings 30 dB
of improvement in ADC linearity with respect to reference-induced distortion. It should be
remarked here that the accuracy of this calibration scheme is limited by the quantization
effects of the back-end that processes the distance estimates required for the algorithm.

Iniyavan Elumalai Master of Science Thesis


5-5 Summary 69

−20 SNDR = 84.4 dB

−40
Amplitude [dB]

SFDR = 84.43 dB

−60

−80

−100

−120
0 10 20 30 40 50
Frequency [MHz]

Figure 5-23: Coarse ADC performance with Cref = 20Cdac after calibration

5-5 Summary

This chapter presented the implementation of the proposed reference scheme in the coarse
stage of the prototype ADC. A digital calibration algorithm was proposed to correct for the
ADC non-linearity caused by errors in Vref stored on a stand-alone capacitor. Finally, the
simulation results before and after calibration were presented, showing that 84 dB of linearity
was achieved with respect to reference-induced distortion.

Master of Science Thesis Iniyavan Elumalai


70 Digital Correction of Reference Errors

Iniyavan Elumalai Master of Science Thesis


Chapter 6

Conclusion

A novel reference scheme suitable for high-end SAR converters has been proposed in this dis-
sertation. In the case of a SAR converter with conventional reference circuitry, the conversion
speed is determined by how fast the reference buffer can settle the transients induced by DAC
switching during every bit cycle. With a large number of bit cycles and capacitive arrays in
the case of high-resolution converters, the reference buffer becomes the most power-hungry
block in order to provide very low output impedance. This essentially limits the SAR con-
verters from achieving good power efficiency at higher speeds. The proposed reference scheme
eliminates this speed limit by employing a stand-alone capacitor to hold the reference voltage
during SAR conversion. Reference settling is limited only by switch resistance and intercon-
nect parasitics. As a result, bit cycles become significantly shorter, which help accommodate
a large number of bit cycles in shorter conversion cycles.
When the stand-alone reference capacitor is valued moderately for area concerns, the
code-dependent errors induced in the reference voltage due to DAC switching become signifi-
cant, leading to heavy distortion. Two different, elaborate calibration techniques are proposed
that help remove the reference-induced linearity. As a proof-of-concept, the proposed refer-
ence scheme is implemented in a 12-bit, 100 MS/s, two-stage, SAR-assisted pipeline ADC in
28 nm CMOS. With both the stages in the pipeline employing moderately-valued, stand-alone
reference capacitors, the calibration techniques help to suppress the thus induced distortion
to less than 84 dB, effectively not degrading the SNDR.
The prototype ADC, designed by the research team in Broadcom Netherlands, along
with the proposed reference scheme consumes a total of 1.81 mW. Twenty-five percent of this
power (460 uW) is spent in pre-charging the reference capacitors during the Sample phase.
With the near-Nyquist SNDR at 68 dB, the Schreier FoM achieved is 175.5 dB. Figure 6-1
shows the reported ADC FoM in the context of the state-of-the-art SAR ADCs (from recent
ISSCC and VLSI publications). With the aid of the implemented reference scheme, this work
clearly pushes the power envelope at moderate speeds. This can serve as an excellent starting
point for time-interleaved architectures, with which we can move the energy efficiency to
higher speeds, ultimately achieving the target FoM for simultaneous full-band capture in a
mobile radio front end.

Master of Science Thesis 71 Iniyavan Elumalai


72 Conclusion

190
Target
180

Prototype ADC
170 with the implemented reference scheme

160
FOMS [dB]

150

140

130 SAR
SAR-TI
120

110
1.E+04 1.E+05 1.E+06 1.E+07 1.E+08 1.E+09 1.E+10 1.E+11
f s [Hz]

Figure 6-1: Schreier FoM plot of the state-of-the-art ADCs along with the prototype

6-1 Thesis Contributions

Some of the salient features of the implemented reference scheme are summarized in this
section.

Drastic reduction in bit trial time: By employing a stand-alone capacitor in lieu of


a conventional, active buffer to provide the reference voltage during SAR conversion, the
reference settling time has been improved significantly. This enabled the average bit trial
time to be reduced from 1 ns to 350 ps potentially for the same power consumption. Further
speed improvement is limited by other factors in the SAR loop, namely, the comparator
and the digital logic. If not for them, this usage of reference capacitor would have given an
additional 100 ps benefit per bit cycle.

Charge-based correction of reference-induced non-linearity: In the fine SAR ADC of


the prototype, using a 1 pF reference capacitor gives an THD of 38.9 dB. Since the reference
errors induced can be determined given the knowledge of the switching decisions being made,
they are corrected through a compensation technique in the charge domain. This improves the
THD to 60.5 dB, more than what is necessary for 8-bit linearity in the fine stage. The overhead
is a small bank of compensation capacitors, whose total capacitance is 20 fF, and digital logic
that performs selection from the bank based on switching decisions, which consumes 30 µW
average power.

Iniyavan Elumalai Master of Science Thesis


6-2 Future work 73

Digital correction of reference-induced non-linearity: In the coarse SAR ADC of the


prototype, using a 20 pF reference capacitor gives an THD of 54.6 dB. A digital calibration
algorithm that takes advantage of the pipeline architecture of the prototype is proposed. The
reference errors are estimated and corrected for by processing the residue voltage of the coarse
conversion. This improves the THD to 84.4 dB, essentially not degrading the SNDR of the
whole ADC. The overhead is that the calibration procedure being a deterministic scheme, a
conversion lane is taken off-line and used to measure the estimates required for correction.
A 9-bit, current-steering calibration DAC and digital processing required for calibration also
represent overhead, however, not in terms of power as they are operated once in a second.

Pre-charge buffer with excellent current sourcing capability For pre-charging the ref-
erence capacitors during the Sample phase, a buffer with excellent current sourcing capability
is designed. It achieves very low output impedance with the aid of a negative feedback net-
work that reuses the main bias current. Drive currents as large as 20 times the quiescent
current are obtained.
At this juncture, some remarks have to made about two relevant works in the literature
that also propose replacing the active reference buffer in a SAR converter with a storage
capacitor, or more specifically, a reference in charge domain. Craninckx and Plas [56] proposed
pre-charging a bank of binary-scaled capacitors which will be connected one by one to the
input sampling capacitor, thus adding/subtracting reference charge until the total charge
converges to zero. While they report a good bit trial time of 1 ns, with reference settling
time determined by the charge sharing switches (in 90 nm CMOS), the SNDR achieved
is limited to the 9-bit level. Analog-domain impairments such as switch charge injection,
capacitor parasitics, and more noise sampling events are introduced inside the SAR loop,
which limit linearity. A more recent work by Kapusta, et al. [57] employs a similar pre-
charged reference for the MSB bit trial, where the reference settling is the most stringent.
However, the code-dependence of the reference charge throughout the whole conversion cycle
is not addressed, which implies that they either utilized an auxiliary reference buffer during
the latter bit cycles or that the capacitance is so high that code-dependence is insignificant.
On the other hand, this work stays close to a classical SAR architecture by employing a single
reference source throughout the conversion cycle, and the code-dependence on the reference is
addressed properly through two elaborate compensation techniques that help achieve almost
14-bit linearity.

6-2 Future work

The design presented in this dissertation gives satisfactory results in simulations. Due to
time constraints, the layout of the proposed new blocks has not been done yet. Realizing a
proof-of-concept test chip is the next step. Meanwhile, this section recognizes some of the
possible improvements that can be made to the proposed reference scheme as follows:

• While this work presented effective ways of removing the code-dependence in the ref-
erence caused by DAC switching, it would be better to first investigate the switching
algorithms that cause this code-dependence, and come up with algorithms that present
less dependence themselves. Set-and-down and split-cap algorithms used in this work

Master of Science Thesis Iniyavan Elumalai


74 Conclusion

exhibit quadratic dependence with fairly large standard deviations of the switching
energies versus code. Whereas, some recently proposed algorithms in the literature,
namely, early-reset merged capacitor switching [58], input-range prediction [28], and
charge-average switching [29], present small code-dependence and the resulting distor-
tion induced would be less to start with.

• The residual INL after digital correction of reference errors in the coarse ADC exhibits
a systematic pattern versus code. Even though this is attributed to the fact that the
errors caused in the reference during the additional step have not been accounted for, a
proper analysis of the residual INL pattern has to be performed. This might help predict
the residual pattern through mathematical models and thus achieve better linearity.

• The digital calibration procedure in this work needs external control to ensure collecting
all of the required information. An on-chip digital machine which can control collecting
the calibration data has to be designed.

• The implemented reference scheme involves a single, stand-alone capacitor holding the
reference voltage. This large capacitor has to be replenished during the Sample phase
which consumes a considerable amount of power. In future implementations, more than
one capacitor can be used in a ping-pong fashion. Proper design choices have to be
made after investigating the trade-off between chip area and power consumption.

Iniyavan Elumalai Master of Science Thesis


Appendix A

Code-dependent Charge Steering


from the Reference Source

A DAC switching event in a bit cycle involves switching a capacitor either from a positive to
a negative reference, vice versa or both in a differential manner, as dictated by the switching
decisions and the algorithm being used. This appendix derives a generalized mathematical
model of the charge steered from the reference source during a switching event, and illustrates
the two switching algorithms used in this thesis.
Figure A-1(a) illustrates the switching event for a negative DAC step. Capacitor Csw
is switched from a positive reference (Vref ) to a negative reference (assumed to be ground)
in order to change the DAC voltage from Vdac to Vdac ′ . Capacitors C and C represent the
+ −
capacitances connected to positive and negative references during the switching event. The
corresponding charge steered from the reference source is given by
 ′ 
Q− = −(C+ ) (Vref − Vdac ) − (Vref − Vdac ) . (A-1)
The DAC voltages before and after switching are given by
C+ + Csw
Vdac = Vref , (A-2)
Cdac
and
′ C+
Vdac = Vref , (A-3)
Cdac
where Cdac is the total DAC capacitance (= Csw + C+ + C− ). Equation (A-1) becomes
Csw C+
Q− = . (A-4)
Vref

Similarly, if Csw is switched for a positive step, ie., from ground to Vref , as shown in
Figure A-1(b), the charge required is given by
Csw C−
Q+ = . (A-5)
Vref

Master of Science Thesis 75 Iniyavan Elumalai


76 Code-dependent Charge Steering from the Reference Source

Vref Vref

C+ Csw C+
Csw C+
Q− = Cdac Vref

C− C− Csw

(a) Negative DAC switching

Vref Vref

C+ C+ Csw
Csw C−
Q+ = Cdac Vref

C− Csw C−

(b) Positive DAC switching

Figure A-1: Positive and negative DAC steps

Based on the switching algorithm, either or both of Equations (A-4) and (A-5) apply
during a switching event. The code-dependence part of the reference charge in a bit cycle
arises from the different possible combinations of C+ and C− .

A-1 Set-and-down Switching Algorithm

In set-and-down algorithm [26], all the array capacitors are connected to Vref p at the start
of the Convert phase. During bit cycling, based on the switching decision, a capacitor in
either the positive or the negative DAC is switched to Vref n . Figure A-2 illustrates the DAC
switching events for a positive and a negative switching decision.

As evident, this algorithm always involves a negative DAC step. Equation (A-4) can be
rewritten to derive the reference charge per bit cycle for this algorithm. In an ith bit cycle,
Ci+1 is the switching capacitor (Csw ). The capacitance tied to the positive reference (C+ ) in
a bit cycle is given by the total capacitance of the capacitors that have not been switched yet
i
P
(= Cl ) and those that remain tied to the positive reference after their switching turns.
l=0

Iniyavan Elumalai Master of Science Thesis


A-2 Split-capacitor Switching Algorithm 77

Thus, C+ can be written as


i
X X
C+ = Cl + s, (A-6)
l=0 s∈Sd
where Sd ⊂ {Cm |(i + 1) < m < N }.
The choice of subset Sd is code-dependent. Thus, the total charge required in each bit cycle
is given by
!
i
P P
Ci+1 Cl + s
l=0 s∈Sd
Qi = (Vref p − Vref n ) (A-7)
Cdac
for i = N − 2, N − 3, ..., 0.

Ci

Di−1 = 1
... ... Vdacp

... ... Vdacn


Vref
Ci
Ci

... ... Vdacp Vref

... ... Vdacn Vref


Ci
Ci

Vref ... ... Vdacp

Di−1 = 0 ... ... Vdacn


Ci

Figure A-2: Illustration of set-and-down switching algorithm in step i

A-2 Split-capacitor Switching Algorithm


In split-capacitor algorithm [24], every capacitor in the array is split into two sub-capacitors.
At the start of the Convert phase, one sub-capacitor in each is connected to Vref p and the other
to Vref n . During bit-cycling, these sub-capacitors are switched differentially as illustrated in
A-3. With this algorithm, every DAC switching event involves both a positive and a negative
DAC step. The total charge required of the reference source in each cycle can be obtained by
rewriting and summing Equations (A-4) and (A-5) accordingly.

Master of Science Thesis Iniyavan Elumalai


78 Code-dependent Charge Steering from the Reference Source

Vref Vref

Ci,1 Ci,2

... ... Vdacp


Di−1 = 1
... ... Vdacn
Vref

Ci,1 Ci,2
Ci,1 Ci,2

... ... Vdacp

... ... Vdacn

Ci,1 Ci,2 Ci,1 Ci,2

... ... Vdacp


Vref
Di−1 = 0
... ... Vdacn

Ci,1 Ci,2

Vref Vref

Figure A-3: Illustration of split-cap switching algorithm in step i

Iniyavan Elumalai Master of Science Thesis


79

Bibliography

[1] J. Mitola, “The software radio architecture,” Communications Magazine, IEEE, vol. 33,
no. 5, pp. 26–38, 1995.

[2] A. Abidi, “The path to the software-defined radio receiver,” Solid-State Circuits, IEEE
Journal of, vol. 42, no. 5, pp. 954–966, 2007.

[3] B. Murmann, “Adc performance survey 1997-2013,” Apr. 2013.

[4] R. Walden, “Analog-to-digital converter survey and analysis,” Selected Areas in Com-
munications, IEEE Journal on, vol. 17, no. 4, pp. 539–550, 1999.

[5] R. Schreier and G. Temes, Understanding Delta-Sigma Converters. Wiley-IEEE Press,


2005.

[6] N. Kurosawa, H. Kobayashi, K. Maruyama, H. Sugawara, and K. Kobayashi, “Explicit


analysis of channel mismatch effects in time-interleaved adc systems,” Circuits and Sys-
tems I: Fundamental Theory and Applications, IEEE Transactions on, vol. 48, no. 3,
pp. 261–271, 2001.

[7] J. McCreary and P. Gray, “All-mos charge redistribution analog-to-digital conversion


techniques. i,” Solid-State Circuits, IEEE Journal of, vol. 10, no. 6, pp. 371–379, 1975.

[8] Y. Zhu, C.-H. Chan, U.-F. Chio, S.-W. Sin, S.-P. U, R. Martins, and F. Maloberti, “A 10-
bit 100-ms/s reference-free sar adc in 90 nm cmos,” Solid-State Circuits, IEEE Journal
of, vol. 45, no. 6, pp. 1111–1121, 2010.

[9] Z. Cao, S. Yan, and Y. Li, “A 32 mw 1.25 gs/s 6b 2b/step sar adc in 0.13 um cmos,”
Solid-State Circuits, IEEE Journal of, vol. 44, no. 3, pp. 862–873, 2009.

[10] K. Bult, “The effect of technology scaling on power dissipation in analog circuits,” in
Analog Circuit Design, pp. 251–294, Springer, 2006.

[11] W. R. Bennett, “Spectra of Quantized Signals,” 1948.

Master of Science Thesis Iniyavan Elumalai


80 Bibliography

[12] F. Kuttner, “A 1.2v 10b 20msample/s non-binary successive approximation adc in 0.13
um cmos,” in Solid-State Circuits Conference, 2002. Digest of Technical Papers. ISSCC.
2002 IEEE International, vol. 2, pp. 136–137, 2002.

[13] B. Razavi, Design of Analog CMOS Integrated Circuits. New York, NY, USA: McGraw-
Hill, Inc., 1 ed., 2001.

[14] S. Jamal, D. Fu, N.-J. Chang, P. Hurst, and S. Lewis, “A 10-b 120-msample/s time-
interleaved analog-to-digital converter with digital background calibration,” Solid-State
Circuits, IEEE Journal of, vol. 37, no. 12, pp. 1618–1627, 2002.

[15] S. Louwsma, A. J. M. Van Tuijl, M. Vertregt, and B. Nauta, “A 1.35 gs/s, 10 b, 175 mw
time-interleaved ad converter in 0.13 um cmos,” Solid-State Circuits, IEEE Journal of,
vol. 43, no. 4, pp. 778–786, 2008.

[16] D. Stepanovic, Calibration Techniques for Time-Interleaved SAR A/D Converters. PhD
thesis, EECS Department, University of California, Berkeley, Dec 2012.

[17] K. Doris, E. Janssen, C. Nani, A. Zanikopoulos, and G. Van Der Weide, “A 480mw
2.6gs/s 10b 65nm cmos time-interleaved adc with 48.5db sndr up to nyquist,” in Solid-
State Circuits Conference Digest of Technical Papers (ISSCC), 2011 IEEE International,
pp. 180–182, 2011.

[18] H. J. M. Veendrick, “The behaviour of flip-flops used as synchronizers and prediction


of their failure rate,” Solid-State Circuits, IEEE Journal of, vol. 15, no. 2, pp. 169–176,
1980.

[19] S.-W. Chen and R. Brodersen, “A 6-bit 600-ms/s 5.3-mw asynchronous adc in 0.13-
cmos,” Solid-State Circuits, IEEE Journal of, vol. 41, no. 12, pp. 2669–2680, 2006.

[20] T. Tong, “Design techniques for successive approximation register analog-to-digital con-
verters,” 2011.

[21] C. Lee and M. Flynn, “A sar-assisted two-stage pipeline adc,” Solid-State Circuits, IEEE
Journal of, vol. 46, no. 4, pp. 859–869, 2011.

[22] Y.-D. Jeon, Y.-K. Cho, J.-W. Nam, K.-D. Kim, W.-Y. Lee, K.-T. Hong, and J.-K. Kwon,
“A 9.15mw 0.22mm2 10b 204ms/s pipelined sar adc in 65nm cmos,” in Custom Integrated
Circuits Conference (CICC), 2010 IEEE, pp. 1–4, 2010.

[23] B. Verbruggen, M. Iriguchi, and J. Craninckx, “A 1.7mw 11b 250ms/s 2x interleaved


fully dynamic pipelined sar adc in 40nm digital cmos,” in Solid-State Circuits Conference
Digest of Technical Papers (ISSCC), 2012 IEEE International, pp. 466–468, 2012.

[24] B. Ginsburg and A. Chandrakasan, “An energy-efficient charge recycling approach for a
sar converter with capacitive dac,” in Circuits and Systems, 2005. ISCAS 2005. IEEE
International Symposium on, pp. 184–187 Vol. 1, 2005.

[25] Y.-K. Chang, C.-S. Wang, and C.-K. Wang, “A 8-bit 500-ks/s low power sar adc for
bio-medical applications,” in Solid-State Circuits Conference, 2007. ASSCC ’07. IEEE
Asian, pp. 228–231, 2007.

Iniyavan Elumalai Master of Science Thesis


81

[26] C.-C. Liu, S.-J. Chang, G.-Y. Huang, and Y.-Z. Lin, “A 10-bit 50-ms/s sar adc with a
monotonic capacitor switching procedure,” Solid-State Circuits, IEEE Journal of, vol. 45,
no. 4, pp. 731–740, 2010.
[27] V. Hariprasath, J. Guerber, S.-H. Lee, and U.-K. Moon, “Merged capacitor switching
based sar adc with highest switching energy-efficiency,” Electronics Letters, vol. 46, no. 9,
pp. 620–621, 2010.
[28] H.-Y. Huang, J.-Y. Lin, C.-C. Hsieh, W.-H. Chang, H.-H. Tsai, and C.-F. Chiu, “A 9.2b
47fj/conversion-step asynchronous sar adc with input range prediction dac switching,” in
Circuits and Systems (ISCAS), 2012 IEEE International Symposium on, pp. 2353–2356,
2012.
[29] C.-Y. Liou and C.-C. Hsieh, “A 2.4-to-5.2fj/conversion-step 10b 0.5-to-4ms/s sar adc
with charge-average switching dac in 90nm cmos,” in Solid-State Circuits Conference
Digest of Technical Papers (ISSCC), 2013 IEEE International, pp. 280–281, 2013.
[30] B. Ginsburg and A. Chandrakasan, “500-ms/s 5-bit adc in 65-nm cmos with split ca-
pacitor array dac,” Solid-State Circuits, IEEE Journal of, vol. 42, no. 4, pp. 739–747,
2007.
[31] M. Saberi, R. Lotfi, K. Mafinezhad, and W. Serdijn, “Analysis of power consumption
and linearity in capacitive digital-to-analog converters used in successive approximation
adcs,” Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 58, no. 8,
pp. 1736–1748, 2011.
[32] W. Liu and Y. Chiu, “An equalization-based adaptive digital background calibration
technique for successive approximation analog-to-digital converters,” in ASIC, 2007.
ASICON ’07. 7th International Conference on, pp. 289–292, 2007.
[33] W. Liu, Y. Chang, S.-K. Hsien, B.-W. Chen, Y.-P. Lee, W.-T. Chen, T.-Y. Yang, G.-
K. Ma, and Y. Chiu, “A 600ms/s 30mw 0.13 um cmos adc array achieving over 60db
sfdr with adaptive digital equalization,” in Solid-State Circuits Conference - Digest of
Technical Papers, 2009. ISSCC 2009. IEEE International, pp. 82–83,83a, 2009.
[34] W. Liu, P. Huang, and Y. Chiu, “A 12b 22.5/45ms/s 3.0mw 0.059mm2 cmos sar adc
achieving over 90db sfdr,” in Solid-State Circuits Conference Digest of Technical Papers
(ISSCC), 2010 IEEE International, pp. 380–381, 2010.
[35] R. Xu, B. Liu, and J. Yuan, “Digitally calibrated 768-ks/s 10-b minimum-size sar adc
array with dithering,” Solid-State Circuits, IEEE Journal of, vol. 47, no. 9, pp. 2129–
2140, 2012.
[36] G. Wegmann, E. Vittoz, and F. Rahali, “Charge injection in analog mos switches,”
Solid-State Circuits, IEEE Journal of, vol. 22, no. 6, pp. 1091–1097, 1987.
[37] H.-S. Lee, D. Hodges, and P. Gray, “A self-calibrating 15 bit cmos a/d converter,” Solid-
State Circuits, IEEE Journal of, vol. 19, no. 6, pp. 813–819, 1984.
[38] B. Murmann and B. Boser, “A 12 b 75 ms/s pipelined adc using open-loop residue
amplification,” in Solid-State Circuits Conference, 2003. Digest of Technical Papers.
ISSCC. 2003 IEEE International, pp. 328–497 vol.1, 2003.

Master of Science Thesis Iniyavan Elumalai


82 Bibliography

[39] C. Grace, P. Hurst, and S. Lewis, “A 12-bit 80-msample/s pipelined adc with boot-
strapped digital calibration,” Solid-State Circuits, IEEE Journal of, vol. 40, no. 5,
pp. 1038–1046, 2005.

[40] J. Keane, P. Hurst, and S. Lewis, “Background interstage gain calibration technique for
pipelined adcs,” Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 52,
no. 1, pp. 32–43, 2005.

[41] A. Panigada and I. Galton, “Digital background correction of harmonic distortion in


pipelined adcs,” Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 53,
no. 9, pp. 1885–1895, 2006.

[42] B. Sahoo and B. Razavi, “A 12-bit 200-mhz cmos adc,” Solid-State Circuits, IEEE Jour-
nal of, vol. 44, no. 9, pp. 2366–2380, 2009.

[43] N. Sun, “Exploiting process variation and noise in comparators to calibrate interstage
gain nonlinearity in pipelined adcs,” Circuits and Systems I: Regular Papers, IEEE
Transactions on, vol. 59, no. 4, pp. 685–695, 2012.

[44] A. Karanicolas, H.-S. Lee, and K. Barcrania, “A 15-b 1-msample/s digitally self-
calibrated pipeline adc,” Solid-State Circuits, IEEE Journal of, vol. 28, no. 12, pp. 1207–
1215, 1993.

[45] I. Galton, “Digital cancellation of d/a converter noise in pipelined a/d converters,”
Circuits and Systems II: Analog and Digital Signal Processing, IEEE Transactions on,
vol. 47, no. 3, pp. 185–196, 2000.

[46] H.-C. Liu, Z.-M. Lee, and J.-T. Wu, “A 15-b 40-ms/s cmos pipelined analog-to-digital
converter with digital background calibration,” Solid-State Circuits, IEEE Journal of,
vol. 40, no. 5, pp. 1047–1056, 2005.

[47] Y.-S. Shu and B.-S. Song, “A 15-bit linear 20-ms/s pipelined adc digitally calibrated
with signal-dependent dithering,” Solid-State Circuits, IEEE Journal of, vol. 43, no. 2,
pp. 342–350, 2008.

[48] W. M. C. Sansen, Analog Design Essentials (The International Series in Engineering


and Computer Science). Secaucus, NJ, USA: Springer-Verlag New York, Inc., 2006.

[49] J. McNeill, M. Coln, and B. Larivee, “"split adc" architecture for deterministic digital
background calibration of a 16-bit 1-ms/s adc,” Solid-State Circuits, IEEE Journal of,
vol. 40, no. 12, pp. 2437–2445, 2005.

[50] H.-S. Lee, “A 12 bit 600 ks/s digitally self-calibrated pipeline algorithmic adc,” in VLSI
Circuits, 1993. Digest of Technical Papers. 1993 Symposium on, pp. 121–122, 1993.

[51] J. Ingino, J. and B. Wooley, “A continuously-calibrated 10 m sample/s 12 b 3.3 v adc,”


in Solid-State Circuits Conference, 1998. Digest of Technical Papers. 1998 IEEE Inter-
national, pp. 144–145, 1998.

[52] P. R. Gray, Analysis and Design of Analog Integrated Circuits. Wiley Publishing, 5th ed.,
2009.

Iniyavan Elumalai Master of Science Thesis


83

[53] R. Carvajal, J. Ramirez-Angulo, A. Lopez-Martin, A. Torralba, J. Galan, A. Carlosena,


and F. Chavero, “The flipped voltage follower: a useful cell for low-voltage low-power
circuit design,” Circuits and Systems I: Regular Papers, IEEE Transactions on, vol. 52,
no. 7, pp. 1276–1291, 2005.

[54] C.-H. Lin and K. Bult, “A 10-b, 500-msample/s cmos dac in 0.6 mm2,” Solid-State
Circuits, IEEE Journal of, vol. 33, no. 12, pp. 1948–1958, 1998.

[55] M. Pelgrom, A. C. J. Duinmaijer, and A. Welbers, “Matching properties of mos transis-


tors,” Solid-State Circuits, IEEE Journal of, vol. 24, no. 5, pp. 1433–1439, 1989.

[56] J. Craninckx and G. Van der Plas, “A 65fj/conversion-step 0-to-50ms/s 0-to-0.7mw 9b


charge-sharing sar adc in 90nm digital cmos,” in Solid-State Circuits Conference, 2007.
ISSCC 2007. Digest of Technical Papers. IEEE International, pp. 246–600, 2007.

[57] R. Kapusta, J. Shen, S. Decker, H. Li, and E. Ibaragi, “A 14b 80ms/s sar adc with
73.6db sndr in 65nm cmos,” in Solid-State Circuits Conference Digest of Technical Papers
(ISSCC), 2013 IEEE International, pp. 472–473, 2013.

[58] J. Guerber, H. Venkatram, T. Oh, and U.-K. Moon, “Enhanced sar adc energy efficiency
from the early reset merged capacitor switching algorithm,” in Circuits and Systems
(ISCAS), 2012 IEEE International Symposium on, pp. 2361–2364, 2012.

Master of Science Thesis Iniyavan Elumalai


84 Bibliography

Iniyavan Elumalai Master of Science Thesis

You might also like