You are on page 1of 13

Embedded Analog-to-Digital Converters

Klaas Bult
Broadcom Corporation
Bunnik, The Netherlands
bult@broadcom .com

Abstract - Systems-on-Chips (SoCs) have Signal Processors (DSP). Examples can be found in
become a reality in the past decade. Several dozens consumer applications like Cellular phones, DVD play-
of different functional blocks are being integrated on ers, Multi-media players and Set-Top boxes.
a single die, reaching transistor counts of up to half a As shown in the typical example of fig. 1, these
billion. From the Analog portion of an SoC the Data systems usually also contain multiple Analog Front-
Converters are probably among the most challeng- Ends (AFE), Analog Back-Ends (ABE), as well as RF
ing blocks, often limiting system performance and Receive and Transmit functions. It is not uncommon
dominating power dissipation. However, require- that the number of Analog/RF functions exceeds a
ments regarding yield, die-size, scalability, noise dozen or more.
immunity, power and the fact that logic is almost for
free, cause distinct differences between embedded B. A typical Analog Front-End (AFE)
Data Converters and their stand-alone, usually gen- Fig. 2 shows a typical AFE. With the idea in
eral purpose, counterparts. This paper describes mind that digital is (almost) for free, the trend is to keep
these differences and provides an overview of the the analog circuitry to a minimum, pushing to the digital
state-of-the art in Analog-to-Digital Conversion. domain whatever is possible. Starting from the back, the
minimum required function is the ADC. Usually a
I. INTRODUCTION Track& Hold or a Sample & Hold Amplifier (SHA) is
preceding the ADC. For anti-aliasing or noise reduction
In a world "going digital" one of the dominant purposes some form of Low-Pass Filtering (LPF) is
remaining Analog blocks is the Data Converter. With implemented , although often just the natural poles of the
the cost of digital functionality going down exponen- other blocks are deemed sufficient. Active filtering is
tially along the path of technology scaling, trading ana- usually avoided as it is often simpler (as well as lower
log circuitry for digital becomes more and more power and cost) to increase the sampling speed a bit to
attractive. The Data Converter is the minimum required allow simple first-order filtering. A Programmable-Gain
Analog function, often limiting overall system perfor- Amplifier (PGA) adjusts the incoming signal to opti-
mance. Within Analog Front-Ends (AFE) or RF-Receiv- mally fit to the reference voltage of the ADC. Omitting
ers (RF-RX) the Analog-to-Digital Converter is one of this function and absorbing this in the dynamic range of
the hardest blocks to design as well. Being embedded the ADC is usually too costly in terms of power dissipa-
imposes extra problems on top of that. Limiting the tion.
scope of the discussion to Nyquist-Rate (High Speed)
Analog-to-Digital Converters, this paper discusses what
it means to be "embedded" (section II, III and IV), the
difference between embedded and stand-alone data con-
I
verters from a system perspective (section V), the state- processor
of-the-art in Analog-to-Digital Conversion (section VI,

.-
,

VII and VIII), how to choose an optimal architecture


(section IX) as well as the future of embedded design
(section X).

II. SYSTEM-ON-A-CHIP (SoC) II.


A. A typical SoC
Nowadays complete and very complex systems
are integrated on one single die and for the larger sys-
tems the number of transistors is approaching a billion.
By far the largest portion of that is in the digital part of
the system, which usually contains Multi-core GHz Pro-

Fig. 1. A typical System-on-a-Chip (SoC). Often multiple
Analog Front-Ends (AFE), Analog Back-Ends (ABE), RF-
Receive (RF-RX) and RF-Transmit (RF-TX) blocks are
cessors, multiple Mbytes of memory, various Media integrated along with multiple DSPs, memory and proces-
Access Controllers (MAC) and several dedicated Digital sor blocks .

978-1-4 244-4353-6 /09 /$ 25 .00 ©2 0 0 9 IEEE

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
Fig. 2. A typical Analog Front-End (AFE) of an
SoC: Programmable Gain Amplifier (PGA), Low-
Pass Filter (LPF), Sample & Hold Amplifier
(SHA) and Analog-to-Digital Converter (ADC).
~~ TUNER

Fig. 3. A typical SoC RF Front-End (RFFE). Most Radio/Tuner


architectures are either Direct-Conversion or Low-IF.

C. A typical RF Receiver (RF-RX) C. Cross- Talk


Similarly to AFE blocks, a typical RF Receiver For a long time cross-talk was believed to be a
(RF-RX), as shown in fig. 3, is also kept as simple as major road-block towards the realization of SoCs: a bil-
possible. Whether Narrowband Radio or Wideband lion digital transistors generate a lot of noise through the
Tuner, the architecture of choice is either Direct Conver- substrate and pin-connections, forcing Analog designers
sion or Low-IF, with significant support of digital func- to design very cross-talk insensitive circuits. But as was
tionality to reduce the effect of circuit imperfections in pointed out in [I] , separating the analog functions on a
the Analog/RF domain. The huge gain necessary in the different die also suffers from X-talk problems, in that
receive chain is obtained through Variable Gain Ampli- case primarily from the digital output pins that form the
fiers (VGA) distributed along the signal path to opti- interface with the fully digital die. Over time however,
mize dynamic Range of the receiver and the ADC. analog designers have developed a way to cope with
that, by choosing high-Ohmic substrates, using guard-
III. WHAT DOES "EMBEDDED" MEAN? rings, using distance, choosing fully differential circuits
A. Process-choice and Process-Options and choosing pin-positions very carefully. At this
As pointed out in the previous section, nowadays moment, although care is always needed, cross-talk is
complete systems are integrated on one single die occu- no longer in the top-three list-of-problems of embedded
pying as much as hundreds of square millimeters . By far Data Converter design.
the largest portion of that is in the digital part of the sys- D. Testing
tem. Although it's hard to give a generic number for the As opposed to their Stand-Alone counterparts,
percentage that Analog circuitry occupies in an SoC, it embedded Data Converters normally don't have their
is fair to say that in the vast majority of SoC chips the digital I/O's connected to external pins. Digital signals
area is dominated by digital circuitry. Therefore the pro- either come from (in the case of the DAC) or go to (in
cess-choice of an SoC is usually completely dominated the case of the ADC) the digital part on the same die.
by digital requirements and the requirements for Analog This not only saves significant amounts of power, but
design playa very small role. This makes perfect sense also relieves the design from the burden of bringing pos-
since with a billion digital transistors on a chip, the pro- sibly very high-frequency digital signals on or off the
cess should be optimized for digital design to optimize chip, along with the associated problems of cross-talk .
the overall yield and cost. For a 12-bit ADC or DAC, the isolation between the
For the same reason, "Analog process options", Most Significant Bit (MSB) and the analog signal has to
like MiM caps or double poly are not allowed. It cannot be better than SOdB, especially at high frequencies a
be justified to increase the total chip-cost by 5% for an very difficult task indeed.
Analog option that is not being used by say 95% of the For testing purposes however, the digital signals
chip area. Analog designers just have to live with that have to be routable to test pins and the problem could
environment and have to make their circuits work with rise again if the sampling speed is high enough. But suit-
what the digitally optimized process offers them. able solutions have been found to address that problem.
B. Yield In the case of the ADC, memory can be placed
In order to achieve a good overall chip yield, on chip to capture a certain amount of data, that subse-
Analog circuitry needs to have a very high yield as well. quently can be read out at a much slower pace, through
The large ratio between Analog and Digital area ampli- test-pins . This can reduce the problem to levels below
fies the significance of Analog yield: it is very costly if a the noise-floor of the design. The required digital mem-
150mm2 die has to be thrown away because a O.2mm2 ory usually is available in the digital system anyway.
analog function has a yield problem. In fact, Analog cir- For DACs the situation is even nicer. In many
cuitry needs to yield close to defect-density to be cases a Direct Digital Frequency Synthesizer (DDFS) is
acceptable. This is quite different from the situation for available to generate the tones needed for testing. Often
stand-alone designs. even a dual DDFS is used to generate the tones for two-

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
tone testing. The only interfaces with the outside world B. Fast Porting and size ofAnalog Organization
are the test-pins to select the exact frequency (often with Unfortunately, the huge portfolio of IP-blocks
24-bit accuracy), the test-pins to select the amplitude of becomes obsolete every 18 to 24 month, when the next
the digital signals and the Analog output signal. All that CMOS-node is becoming economically viable. So,
is needed is a Spectrum Analyzer to analyze the output being embedded not only means that you need to be able
signal. In the case a DDFS is not already available in the to port a single design quickly and often from one prod-
digital system, that function is often integrated addition- uct generation to another (with often increased specs),
ally for test-purposes only. In modem CMOS processes but you actually need to port the entire portfolio from
the size of a DDFS is so small that cost is no longer a one technology to the next every 18 to 24 month. As a
reason not to do this. result it is prudent to look ahead to whatever the next
Most SoCs incorporate a receive- as well as a technology looks like and check whether your design
transmit-function, which means that both a DAC as well will also work (with possibly some minor changes) in
as an ADC are present. This enables Analog loop-back the next process node. Doing something that only works
in which the DAC drives the ADC, allowing the imple- in the current technology means postponing challenges
mentation of a fully digital test as a very powerful and to the future. Because digital circuits are easily ported
cost-effective test-scenario. Examples can be found in from one technology to the next, to be able to keep up
Ethernet, DSL and SerDes. with their Tape-Out schedules, a significant size of orga-
nization supporting the Analog/RF circuits is needed.
E. Digital Signal Processing Power
So far in this section mostly additional demands c. System Knowledge
on Analog circuit design imposed by the integration A lot of system knowledge usually is available in
onto an SoC have been discussed. However, being an SoC design environment. Dedicated Analog design
embedded within large amounts of digital circuitry also can take advantage of this specific system knowledge.
means that a lot of digital processing power is available, This also means that in most cases the Analog design
which can be used to the advantage of the Analog will not be general purpose but specifically tailored to
design. Even complex functions like FFTs are available. the system at hand. This can easily lead to seemingly
The general trend towards digitally aided Analog design unbalanced designs where for instance the equivalent
(where techniques like digital calibration and error-cor- ENOB from a noise point of view is 1 or 2 bits higher
rection techniques are fully exploited) allows for better than the equivalent ENOB for distortion, simply
Analog performance or for reduced power, while main- because that is the best choice from a system perspec-
taining the same performance. tive (see section V.G).
Contrary to what might be expected, the vast
IV. DEMANDS ON THE ORGANIZATION amount of system knowledge usually leads to continu-
As stated previously, large SoC's contain dozens ously changing analog specs. The constant quest for the
of sub-systems each requiring their own Analog/RF optimal system performance includes the optimization
blocks. Systems of this size always require the latest of the specifications for the Analog blocks, making
CMOS node in order to hit the cost-target for the end- Analog design specifications a constantly moving tar-
applications. Often the integration-level of these sys- get. A significant amount of systems knowledge is
tems are limited by yield issues and higher integration- therefore needed from Analog designers in order to be
levels require the higher packing density of the next able to negotiate the most appropriate target specifica-
technology-node. Supporting the Analog and RF tions for the Analog blocks.
requirements of these systems puts very specific D. Programmable Design and Design for Bug-Fixing
demands on the organization. Modem CMOS processes are extremely expen-
A. Size ofthe Portfolio sive and dedicated test-chips are often very hard to jus-
Dozens of systems cannot be designed from tify. Even Multi-Project Wafers (MPW) are becoming
scratch and expected to be successfully integrated on an scarce. New designs are often expected to go in produc-
SoC. Typically 90% - 95% of the blocks have to be tion in one single pass with only metal-spins to solve
ready on the shelf and tested in previous systems. Only a occasional bugs. This requires not only very careful
few blocks per generation can be added if any reason- design, but it is also prudent to add a significant amount
able rate of success is expected. This requires a very of dummy transistors (fingers) and route all the connec-
large portfolio of sub-systems, including their Analogi tions in metal. On top of that, many bias-currents, refer-
RF blocks, to be production-tested and ready. For that ence-voltages, load-resistors, filtering-capacitors, etc.
reason it has become very hard for smaller companies to can be made programmable, to accommodate unex-
enter this business space. pected parameter skews.

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
V SYSTEM ASPECTS frequency. For reasons like anti-aliasing or noise sub-
sampling and down-folding it is often preferable to have
One of the more subtle differences between some margin between the highest signal frequency and
Stand-Alone and Embedded Data Converters is that the the Nyquist frequency. That means that often the perfor-
latter can be even more specifically tailored to the needs mance of the data converter does not have to be flat up
of the system, as is shown below. to Nyquist. If that is the case it would be overkill and a
waste of power not to take advantage of that in the
A. SNR versus THD design.
SNR and THD both have an equal effect on the
ENOB of a Data Converter. For that reason Stand-Alone D. Latency
ADCs usually have balanced specification with respects Latency can be a problem in some systems. Usu-
to noise and distortion. Some systems however, are ally that depends on whether or not a loop is fed back
more sensitive to nonlinearities, like narrow-band RF- around the data converter. If there is, then the stability of
systems, where strong blockers in adjacent channels can the loop will be negatively influenced by an increase in
leak in to the wanted channel through inter-modulation latency. Often the latency in the digital domain is much
distortion and wipe out the wanted channel. bigger than the latency of the data converter and this
Other systems are often more sensitive to noise. becomes part of the negotiation with the system/digital
An example is the quantization of a wide-band signal. designers. For this reason, pipeline converters may have
Due to the broadband nature of the signal, the statistical to be excluded as a result of a strong restriction on
properties start resembling that of a Gaussian-distrib- latency.
uted signal. A well known aspect of that is that,
although So events do happen, the signal spends most of E. Bit Error Rate Performance
its time between +/- 2a. In systems with some amount Bit Error Rate (BER) is a design parameter which
of error-correction, the influence of errors caused by is often not discussed in ADC presentations or papers. It
rare So events is so small that the Data Converter does is not simple to deal with in simulation and very time-
not have to be able to handle such large signals without consuming during characterization. However, especially
some form of performance degradation. That means that in fast-sampling data converters, BER can be a very
at full-scale, the THD does not necessarily have to meet dominant factor in design, dependent on the system
the same value as the SNR. In fact the "imbalance" can requirements. As an example, Cable applications like
be significant, like 1 or 2 bits. Set-Top Boxes or Cable Modems require BERs in the
Data Converters with more stringent specifica- order of 10-12 and Ethernet applications require even
tions on SNR than on THD may seem unbalanced (and lower BERs of around 10-15 . Such demanding specifica-
are much harder to publish for sure). But on the other tions can have a strong effect on the choice of architec-
hand, that might be exactly what the system is asking for ture. Fast sampling applications with stringent
and the Converter design should take advantage of that. requirements on BER may require either Flash- or Fold-
ing-architectures to meet the BER specifications (see
B. Quantization Noise also sections VI.A and VI.F). However, BER specifica-
Quantization noise seems like a straightforward tions can vary widely from system to system. Measure-
aspect of a converter and often does not get much atten- ment systems, in which the final result may be averaged
tion. If one needs 9.5 ENOB, then a 10-bit should do the over time can be more relaxed in their BER-require-
job. However, an ideal10-bit (without any noise, distor- ments. A BER of 10-6 will not seriously degrade the
tion or jitter) is already limited to 10.0 ENOB. In order SNR of oscilloscope front-ends, even if every single bit-
to achieve the required 9.5 ENOB all the other effects error is a full-scale error. Many wireless systems aren't
together (noise, jitter and distortion) should also achieve even requiring anything more stringent than a BER of
10.0 ENOB in order to obtain 9.5 ENOB for the total 10-4 .
design. By deciding to go to 11 or 12 bits somewhat
more room for the other effects is obtained in a fairly F Accuracy ofReferences and Offsets
cheap way (see section V.I). Adding an additional least- The accuracy requirement with respect to refer-
significant bit is not costly at all, although the digital ences and offsets can also be very different between
needs to be able to process this. This is a degree of free- embedded converters and their stand-alone counterparts.
dom in embedded design that general purpose convert- The stand-alone version usually has tight specification
ers cannot benefit from. on the accuracy and spread of the reference voltage. The
embedded version has very relaxed specifications for
c. Frequency Response the reference voltages because it is usually preceded by
General purpose ADCs usually are expected to a Programmable Gain Amplifier (PGA) that is automati-
have a flat (or nearly flat) performance up to the Nyquist

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
cally adjusted to allow optimal use of the converters The embedded version uses 45-nm CMOS,
dynamic range. which is required to meet the level of integration and
On the other hand, in the embedded version usu- cost target for the system. The design is operated from a
ally no pins are available for the references, neither to 1.0-V supply but uses thick-oxide (this option is free
use external (precision) components nor for de-cou- because it is required by the digital I/O) for switches,
pling. The problem needs to be solved completely on- allowing rail-to-rail operation. The stand-alone version
chip, whereas the stand-alone version easily can use uses a 250-nm technology because it's less expensive
some pins for this purpose. and uses a 2.5-V supply, the regular supply voltage in
this technology. The MiM-option is used to reduce area
G Embedded versus Stand-Alone ADC for the capacitors used in this design.
To exemplify the differences between embedded The embedded version is realized in 0.5-mm2 .
and stand-alone ADCs discussed in the above, this sec- Area is important because it immediately impacts the
tion compares a typical (but imaginary) stand-alone cost of the SoC. The small feature size and improved
general purpose ADC to a typical (but imaginary) matching of this technology are full exploited. The
embedded ADC. Table 1 shows the specifications of stand-alone version has an area of 8.0-mm 2 . Area is not
both example ADCs. extremely important because the design is pad-limited
Table 1: A typical embedded vs. stand-alone ADC (and the cost per mm 2 is much lower in this relatively
old technology anyway).
Specification Embedded Stand-alone
The embedded version is packaged in a 776-pin
NOB 12 10 BGA which was necessary for the SoC. Pins for the ana-
SQR [dB] 74.0 61.9 log part of the chip (12) are kept to a minimum to reduce
Thermal Noise [dB] 65.1 65.1 the cost. For testing, an additional set of test-pins can be
SNR [dB] 64.6 60.2 configured to allow separate testing of the data con-
THD [dB] 61.1 68.0 verter. The stand-alone version is packaged in a 128-pin
SNDR [dB] 59.5 59.5 EPQFP, where all pins are used for the converter.
ENOB 9.6 9.6
VI. ADC TECHNIQUES
INL(lOb) [lsb] 0.6 0.4
DNL(lOb) [Isb] 0.3 0.3 This section gives an overview of Analog-to-
ERBW[MHz] 80 100 Digital Converter techniques and the state-of-the-art in
Fsamp [MS/s] 320 200 Analog-to-Digital Converters.
CMOS [nm] 45 250 A. Flash-Converters [8-11]
Options Thick-Oxide (free) MiM The Flash-architecture is probably the most
Vsupply [V] 1.0/2.5 2.5 straightforward ADC architecture known. It is widely
Area [mm/] 0.5 8.0 known that Flash-architectures are good for high speed.
Package 776 pin BGA 128-pin EPQFP Partly this is because there is often very little circuitry
Power [mW] 19.6 33.5 between the input signal and the comparator (often just
a simple pre-amp), but probably even more important
In this example, although the total resulting because the comparator decision can be followed by
ENOB is the same, the number of bits in both designs is several digital latches, increasing the BER (see section
different. The embedded design is taking advantage of V.G) dramatically. Architectures that require a decision
the freedom to choose a few more bits to lower the in one or even in half a clock-cycle, like any subranging
quantization noise. As the application happens to allow architecture (including most Two-Step and Pipeline
lower THD (because the input is a broadband signal architectures) do not have this option (see section VI.B).
with an amplitude distribution close to Gaussian) the The Flash-architecture remains the architecture of
extra headroom is used for lowering the THD which in choice for the fastest low resolution type of applications
tum allows for lower power, specifically in the continu- [9, 11], often with averaging and interpolation [8, 10]
ous-time part of the AFE. The result of this is an INL/
DNL that is significantly higher in the embedded ver- B. Sub-ranging [12-15]
sion (but on an 12-bit resolution compared to a 10-bit Although the Flash-architecture is the most
resolution of the stand-alone version). In the same way, straightforward and simplest architecture, the number of
the embedded version takes advantage of the fact that comparators and their accuracy in terms of offset grow
only 80-MHz of bandwidth is required, but at the same exponentially with the number of bits required in the
time pushes the clock-frequency to give more headroom converter. The subranging architecture solves that prob-
to filtering. lem by taking a Two-Step or Multi-Step (Pipeline or

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
AiD D/A AiD

Q) Q)

Coarse-bits Fine-bits 0)
c
0)
c
~ ~
Q; Q;
Fig. 4. Subranging architecture. Usually there is 1 bit of a > b >
0 0
overlap between Coarse and Fine ADC to allow correc-
tion of errors of the Coarse-ADC (see fig. 5). Fig. 5. The use of overrange for error correction, a) the
regular situation, b) the signal goes in overrange.
SAR) approach (Fig . 4) . First a Coarse-ADC makes a
rough estimate where the input signal is. Then the archi- extra gain between the stages and in tum the extra gain
tecture zooms in on that estimate, by subtracting the is possible because the next stage does not have to han-
coarse-estimate from the input signal (through the dle the entire range as a result of the zooming-action.
DAC) and subsequently magnifying the residue-signal Almost all architectures using subranging, like Two-
with an amplifier. In a second step a Fine-ADC deter- Step, Subrange, Pipeline and SAR make use of this cor-
mines the value of the input signal more accurately. This rection technique.
idea forms the core of most of the known data-converter D. Pipeline Architecture [16-26]
architectures, including Two-Step, Subranging, Pipelin- Of coarse the principle of sub-ranging can be
ing and Successive Approximation Registers (SAR) repeated in the Fine-ADC. If we cascade several of
based ADCs . those stages , we can perform one sub-range conversion
The action of subranging requires that in one per clock-cycle and each clock-cycle the signal ripples
clock-cycle the Coarse comparators take a decision , through this cascade of sub-range blocks . This architec-
subsequently drive the DAC to subtract the coarse-esti- ture is called a Pipeline ADC and is shown in fig. 6a.
mate from the input sif,rnal and feed the residue-signal to The Pipeline ADC is one of the most popular
an amplifier, which then amplifies the difference and ADC architectures. Extracting only 1 or a few bits per
supplies the result to the Fine-ADC. This requires care- stage (and then passing it on to the next stage, the real
ful timing and precludes the use of digital latches fol- pipelining) keeps the problem per stage simple, the
lowing the comparators to improve the BER. Running throughput high (because only 1 or a few bits are
such an architecture at high clock-speeds requires very resolved per clock-cycle the clock-speed can be rela-
good comparators (very low BERs) or the system tively high) and still reduces the problem for the next
should allow moderate to bad ADC-BERs. stage by the number of bits resolved in the stage (2
times simpler in a l-bit-pcr-stagc architecture). This
C. Error-Correction using Overrange [16]
architecture lends itself well to several forms of error-
If the signal is close to the boundary between two
correction, of which using overrange as discussed above
adjacent sub-ranges, it is possible that the coarse esti-
mate selects the wrong sub-range. However, this prob- is the simplest one. But in the past decade many more
elaborate error-correction techniques have been devel-
lem can be easily solved by reducing the gain of the
oped, allowing either better performance or lower power
residue amplifier and mapping the subranging on a
smaller portion of the input range of the Fine-ADC (fig. for the same performance (see section VII).
5a). The additional part of the Fine-ADC range is called E. Succ essive Approximation Register ADC [27-32]
overrange. If the input signal is close to an edge of the The SAR-ADC uses the idea of subranging in a
subrange and by mistake selects the wrong subrange slightly different way and is shown in fig. 7. Instead of
(fig. 5b), then the residue-signal does not fall in to the using multiple stages, all the bits get resolved in 1 single
intended range of the Fine-ADC, but in the overrange. stage, but at a rate of 1 bit per clock-cycle. The DAC in
However, since the Fine-ADC can still perform well in the feedback-loop slowly approaches the value of the
the overrange, the error can be recovered and the Coarse input signal. To that purpose, the comparator decides
bits can be corrected. each clock-cycle whether the approximation of the DAC
In essence, this technique is possible because the needs to be corrected up or down and the SAR-Iogic
next stage can look more accurately as a result of the

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
a)

AID D/A AID D/A AID D/A


Stage 1 Stage 2 Stage 3

b) Contains
Error-Signal
1 Digital Domain
Fig. 6. A Pipeline Converter with a) the first 3 stages and b) gain-calibration in the digital domain.

keeps track of all the decisions and drives the DAC A significant advantage of Folding is that it does
through a binary-search algorithm . The strong points of not use comparators for the folding-action. As opposed
this architecture are it's simplicity, the total omission of to other subranging architectures the clock-speed does
gain-elements (amplifiers), and the insensitivity to off- not get limited by the fact that comparator decisions
sets. This allows for very low power implementations. have to be used to select the next subrange . In that
However, it needs N clock-cycles for N bits of resolu- respect the Folding-architecture falls in the category of
tion, which limits the maximum achievable conversion Flash-type architectures that in principle can have
rate. Although this architecture has been around for extremely good BER by adding latches at the end. This
quite a long time, the low conversion speed has pushed does increase the latency, but one or two clock-cycles
it's use to the background for a significant amount of does not necessarily have to be a problem .
time. Recently, interleav ing (section VI.G) has been Folding is still being used today in fast sampling
used to solve this issue [27-32], with very good results. ADCs with moderate resolution [34-38]. In [38], folding
Because of it's inherent low-power features and is pushed to another level again, showing a pipelined
progress in interleaving techniques, SAR-ADCs have folding implementation of 729-times folds, with coarse-
become very popular again [27-32]. comparators in every folding-block, making the distinc-
tion between Folding and Pipelining even more subtle .
F. Folding and Interpolating ADC [33-38J
Folding is a technique that has been around for a G Interleaving [29-32, 37-39J
long time and has becom e popular in Bipolar technol- Every ADC-architecture has a certain maximum
ogy [33]. It is a form of Analog pre-processing that aims conversion speed in a certain process technology. When
to reduce the number of comparators by using the same pushed to this limit, the efficiency in terms of perfor-
comparator for more than one decision . In fact, Folding mance-per-mW power-dissipation will go down. If
is an implementation of subranging without making usc either a conversion speed exceeding what was deemed
of a coarse ADC to determine the subrange and not nec- feasible in a particular technology, or improved effi-
essarily using clocked-clements like comparators. ciency is needed, parallelism in the form of interleaving
can be used. The effect of this is an exchange of hard-
ware for speed. Interleaving [39] is implemented with N
slices of the same ADC in parallel, each running at lIN
o times the overall intended speed. When combined in the
J (1)'-
>0)
.- 0
f/)...J
digital domain, the overall intended conversion speed
can be obtained .
:g . There are however, a few problems in this
o ><
o ~ approach that have to be dealt with. When only the ADC
~c..
cnc.. is interleaved and is driven by a single SHA, N switches
« connecting the SHA to each of the N slices are needed .
The SHA supplies each of the slices with samples at the
Fig. 7. Successive Approximation Register (SAR) ADC.

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
rate of that individual slice and is rotating through the N VII. ADVANCED CALIBRATION TECHNIQUES
slices. In that case the SHA itself needs to run at the
overall conversion speed. That makes the SHA the hard- As stated before, the embedded environment is
est block to design and also limiting the conversion ideally suited for digital-aided analog design, allowing
speed. better performance at lower power-dissipation levels.
There are also errors introduced by the interleav- New advances in Error-Correction and Calibration are
ing technique itself. Mismatch between the slices will being discussed in the following paragraphs.
cause offsets and gain-errors for each slice. If N slices A. Gain Calibration Techniques [17-26]
are interleaved and Fsamp is the overall sampling speed, One of the main sources of errors in Pipeline
offsets between the individual slices will cause tones at converters is a deviation of the amplifier gain from the
multiples of FsamplN. If the input frequency equals Fin' ideal value of 2m (see fig. 4). As a result the transition
Gain-errors between the slices will cause tones of Fin from one subrange to another will show a sudden jump,
around the same multiples of FsamplN (at nFsamplN +/- resulting in a spike in the DNL-curve. Many papers
Fin' with 0 < n < N). Generally speaking, these tones have been published to correct these errors [17-26].
will fall in the signal band-of-interest and need to be If an estimate could be made of the actual ampli-
kept to a minimum. Offset and Gain-calibration tech- fier-gain' then these errors could be corrected in the dig-
niques [37] or redundancy [30] have been proposed to ital domain by changing the radix accordingly. The
deal with these problems. principle of estimating the amplifier gain is depicted in
If the SHA cannot meet the required sampling fig. 6b. A small pseudo-random signal is injected at the
speed, it can be built of interleaved slices as well. This input of the DAC in Stage 1 and gets subsequently con-
however, introduces additional errors, usually even verted to the Analog domain. It then goes through the
harder to deal with. Now, due to mismatches, the sam- analog "2x" amplifier and is fed to Stage 2. Simulta-
pling moments might deviate from the ideal uniform neously, in the digital domain, the pseudo-random signal
sampling grid. This also causes tones of Fin around mul- is passed through an ideal "2x" amplifier and fed to the
tiples of FsamplN. These tones however, cannot be can- DAC of Stage 2. If both path match (i.e. if the Analog
celled by the gain-calibration techniques because of "2x" amplifier has a gain of exactly 2), then the two sig-
phase-differences. nals cancel at the input of the "2x" amplifier in stage 2.
Differences in bandwidth also generates tones at Any deviation from the ideal value of 2 will result in a
the same spot in the spectrum, but now the magnitude of finite "error" signal that can be detected in the digital
the tones are signal-frequency dependent. This compli- domain at the output of the ADC of Stage 3 by correlat-
cates the problem very substantially and requires very ing that signal with the pseudo-random signal. This can
significant signal processing to be solved. In [61] an be used to slowly adapt the radix in the digital domain to
embedded solution is shown where all lanes are kept match the Analog gain.
separate in the Analog domain and in the digital domain This technique could be implemented in the fore-
the lanes are combined using a MIMO-FFE (Multiple- ground during start-up, using substantial calibration sig-
Input Multiple-Output Feed Forward Equalizer), allow- nal-levels, but it is also often done in the back-ground
ing the lanes to be equalized individually. This is a per- (during normal operation). The disadvantage is that the
fect example allowed by an embedded environment, pseudo-random signals needed for the training have to
impossible for stand-alone solutions. be small with respect to the regular signal and a correla-
Table 2: Interleaved ADCs pushing the sampling speed tion techniques with long averaging is needed to extract
the error signals correctly, taking a long time to con-
Ref NOB ENOB ERBW F samp N
verge. While during normal operation this might not be
[58] 8 6.5 2.0 20 80 a problem, during production testing this can be very
[59] 11 8.8 0.4 1.0 8 problematic indeed.
[60] 11 9.2 0.4 0.8 4
[29] 10 7.9 1.0 1.35 16
B. Amplifier Nonlinearity Calibration [22,26,53]
Gain-errors are not the only artifacts being cali-
In the past decade interleaving of pipeline ADCs brated in Pipeline ADCs. Recently papers have been
has been pushing the conversion speed in to the GHz published that go much further [22,26, 53].
range. Table 2 gives an overview. Due to the above men- In [53] a converter is proposed that also cali-
tioned additional problems of the interleaving technique brates for 3rd-order amplifier distortion. The principle is
itself, the overall ENOB never reaches the ENOB of a rather simple. Instead of a single pseudo-random signal,
single slice. Note that in table 2 on average the ENOB is a sum of 3 uncorrected pseudo-random signals is
almost 2 bits lower than the NOB. injected at the input of the DAC in the sub-range block
of a Pipeline ADC. If the amplifier has a 3rd order dis-

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
tortion component with coefficient u3' then in the output There are a few great advantages of this tech-
signal of the amplifier a component will be present nique. First of all, from an Analog perspective, no extra
equal to the product of the 3 pseudo-random signals and circuitry and no extra power dissipation is needed, the
u3. This component is then digitized again by the next two halves dissipate exactly 50% of the original power
stage ADC. Correlating the signal at the output of the dissipation. By combining (adding) the signals in the
ADC with the product of the pseudo-random signals digital domain, the original SNR is not degraded as
yields an estimate of u3. This estimate can be used to well. The second great feature of this technique is that
cancel the 3rd-order distortion terms in the digital no calibration signal (like a small pseudo-random sig-
domain. The same paper [53] also shows an implemen- nal) is needed, the circuit can converge during regular
tation of their previously published DAC-DNL cancel- operation and uses the difference between the two
lation. Although active area is relatively high, very good halve-circuits. The problem has become deterministic
results have been obtained indeed. and can converge orders-of-magnitude faster than the
techniques discussed in the previous sections.
C. Split-ADC Calibration Technique [54-57]
To tackle the problem of long convergence times, D. Future Calibration Techniques
techniques have been investigated that make use of In principle, with the vast amounts of signal pro-
deterministic methods as opposed to the statistical meth- cessing power available around embedded Data Con-
ods described in the previous section. Recently a very verter, even more elaborate techniques could be
elegant method for Pipeline-calibration has been pro- considered, for example techniques making use of FFT-
posed [54 - 57]. The technique is based on the simple functions. Of special interest are techniques that make
idea of splitting an ADC in two half-circuits, each con- use of advanced system information like the decision-
suming exactly 50% of the power, with devices exactly point (slicer) SNR. This would allow Data Converter
50% of the original size (in width) and each functioning designs tailored exactly to the needs of the system at
as a complete ADC. If both circuit halves are being hand.
operated simultaneously on the same input signal, and if Currently, most of the advanced calibration tech-
in the digital domain the resulting converted signals are niques are still under investigation at Universities only
added, the overall performance should be exactly equal and implemented in Matlab. It will take a while before
to the performance of the ADC before it was split in to these techniques end up in actual products. But undoubt-
two half-circuits. However, due to random mismatches edly techniques like these will be used in the future,
and spreads, each half ADC will have a different offset allowing even better performance at a lower power
and gain-blocks will have slightly different gains. This level.
can be detected by looking at the differences between
the two circuit halves, any difference between the two VIII. LOW POWER TECHNIQUES
halves has to be an error. The technique is based on
using the difference between the two halves as an input Apart from calibration techniques to lower
to the calibration engine to correct the errors in each of power dissipation (and interleaving which indirectly can
the circuit halves. Since taking the difference between have the same effect), other techniques that aim to
the two halves cancels the regular signal, the calibration reduce power dissipation have been proposed recently.
engine only operates on error signals and converges Most of them try to either reduce the number of
much faster. OpAmps or substantially simplify them.
Systematic errors present in both circuit halves
are not automatically cancelled this way, but with proper A. OpAmp sharing [41-43]
knowledge of the error-mechanisms of a certain archi- A straightforward implementation of a Pipeline
tecture, this may be solved as well. For instance, if this ADC uses one OpAmp per stage. Each OpAmp is oper-
technique is applied to a Pipeline ADC, each half will ated in 2 phases, a reset-phase and an amplify-phase.
produce a typical error plot with jumps at the transition This means that only 50% of the time the OpAmp is
from one subrange to the next. If the two circuit halves actively doing something and the other 50% is wasted.
are deliberately given different coarse decision points, The OpAmp-sharing technique makes use of this situa-
then those jumps will appear at different spots for each tion by using the OpAmp actively in both phases.
half-ADC. The result is that one half-ADC is generating This was first done in Switched-Capacitor cir-
significant errors at a location where the other half-ADC cuits [40] by sharing the OpAmps between two parallel
is hardly making any error at all and vice versa. By circuits that were operated out-of-phase and was called
looking at the difference between the two circuit halves double-sampling. In Pipeline-ADCs this can be done if
these errors can be detected and fed to the error-correc- there are multiple ADCs in parallel, like for instance in
tion block [57]. an I-and Q-path of an IF-section of a receiver, or when
multiple ADCs are interleaved. When this is not the

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
case, then OpAmp-sharing can also be implemented crossing. Comparators are simpler than OpAmps
between successive stages of a pipeline [41-43]. because stability as well as linearity is no issue. This can
Normally, a factor of two in power dissipation not only save a significant amount of power, but may
reduction can be obtained. However, in the case of suc- also enable even lower voltage operation than what is
cessive stages, the stages are usually scaled down in possible for OpAmps. In [47] a lOb 26MS/s pipelined
power to take advantage of reduced SNR requirements ADC is realized in 65nm CMOS, operating from a 1.2V
of the following stage. In that case less than a factor of supply and a power-dissipation of 5.5mW.
two can be gained by this technique.

B. SHA-Iess ADCs [48-52] IX. OPTIMAL ARCHITECTURE?


The Sample & Hold Amplifier (SHA) preceding A. No Real Favorite
most ADCs, has to be able to track the input signal with Architectural choices can have a major impact on
sufficient linearity (similar to the entire ADC) and also the design of circuits. However, many of the architec-
show the same SNR. The OpAmp used in the first pipe- tures discussed above have common properties and dif-
line-stage has to handle the same signal swing and has ferences are not always that obvious. Over the years
similar noise requirements. These blocks are the two almost all architectures have retained their valid spot in
most dominant blocks in linearity and noise and are the spectrum of Data Converters and there are no obvi-
therefore consuming most power. ous winners. Similarly, there are no obvious architec-
For that reason, several proposals have been tural choices for embedded converters.
made to share the OpAmp between these two blocks, When decisions have to be made about architec-
which can lead to significant power-savings. tural choices, comparisons between architectures are
C. Openloop Residue Amplifier and Calibration [22] unavoidable. If data-points are available close to the tar-
With the idea in mind that errors caused by the get specifications, the comparison is easy. However,
residue-amplifier can be corrected by digital calibration often data points are available from (previous) designs
(section VII), [22] have shown that it is feasible to use that are somewhat further away from the target specifi-
openloop amplifiers in stead of closed-loop amplifiers cations and sometime even further away. In those cases
for this purpose. In this particular case a gain of 8x was a fair and relevant comparison can only be made
sufficient and, as a result of the absence of a loop, the through extrapolation or interpolation. The question that
design is completely freed of any stability requirements. needs to be answered is then: given this performance,
That simplifies the amplifier design to a cascoded differ- what does it take to get to the desired performance?
ential pair with resistive loads. An ENOB of 10.7 is Usually the primary aspects being considered are power
achieved while sampling at 75MS/s. and sometimes area. Figures-of-Merit are exactly doing
that, they try to give an answer to the question: which
D. Incomplete Settling and Calibration [44] architecture can do the job for the lowest cost?
Along the same lines, [44] proposes to allow
incomplete settling of its residu amplifiers to obtain B. Figures-of-Merit (FOMs)
higher conversion speeds. Errors caused by this design Popular FOMs [2] are trying to calculate the
are calibrated again. Even lower power than [22] is power it takes per conversion-cycle per lsb and are cal-
shown at similar performance. culated as:

E. Comparator-based ADCs [45-47] FOM = P / (2 ENOB 2F ERBW) (1)


OpAmps are among the most complicated analog in which P = Power Dissipation, ENOB Effective
sub-blocks in data converters. They have to supply suf- Number of Bits and FERBW = Error Resolution Band-
ficient gain to suppress any distortion to a low enough width, which is the frequency at which the performance
level, they have to show low enough noise to support the has dropped by 0.5 ENOB. Usually the minimum of
dynamic range of the converter and they also have to be FERBW and Nyquist-frequency (F clk/2) is used in (1) to
stable in feedback with sometimes varying load-condi- avoid the inclusion of subsampling.
tions. The main task of the OpAmp is to make sure the Sometimes, instead of the Effective Number Of
feedback-capacitors see a sufficiently complete charge- Bits (ENOB), only the number of bits (NOB) are taken,
transfer by ensuring a very low virtual ground level at ignoring any lack in performance drop versus frequency,
the end of the clock-period. which is less useful. Sometimes simply the Nyquist-fre-
Comparator-based ADCs accomplish the same quency is taken instead of FERBW. This ignores any
by using a comparator to switch off a current-source reduced performance below Nyquist. In that case, when
when the charge-transfer is complete. To this purpose applied to oversampling ADCs, FNyquist is normally
the comparator checks the virtual ground for a zero- substituted by 2 times the input bandwidth of the ADC.

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
c. The use ofFOMs the supply voltage (2.5V, 1.8V, 1.2V, 1.0V, 0.9V) was
Even though these FOMs are derived from vast claimed again as the end of the road for SoCs and some
amounts of published data [2, 6], one has to be very companies changed their focus to Multi-Chip Modules
careful to apply this to individual architectures. The data (MCM) for that reason only. In the newest technologies
that was used to derive these FOMs ranged over many (90nm, 65nm, 45nm) the supply voltage hasn't scaled
different architectures, exhibiting vastly different values down nearly as much as was originally predicted [7], but
for ENOB, Fclk , FERBW and Power. On top of that, only new problems came to light, like gate-leakage or low
the best performing converters were used to do the intrinsic gain. Nevertheless, non of these effects so far
curve- fitting. stopped the progress on SoCs and in the past decade
The problem may be best shown using the fol- SoCs have been shipping in all these technologies in
lowing example. Ifa 4-bit Flash ADC (16 pre-amps and many millions per month. It doesn't look like that is
16 comparators) needs to be improved to a 5-bit Flash, going to change for the next technology nodes either.
the power would have to go up 2x for doubling the num- It can been shown that technology scaling has a
ber of pre-amps and comparators and another 4x for strong effect on power dissipation in analog circuits. In
lowering the noise by 1 bit. This would cost 8x in [5] 24 different situations in Analog design have been
power, far more than the 2x predicted by the FOM of studied and categorized. But contrary to popular
equation (1). However, if a designer would keep doing believe, it is not only getting worse with lower supply
this to extend the number of bits to for instance 8, the voltages. Dependent on the what dominates the lower
power dissipation would be horribly high and he would end of the dynamic range (noise or matching) and
be forced to go to another architecture (probably already dependent on whether the specifications are dominated
going from 4 to 5 bits). by small signal, large signal or feedback requirements,
For individual blocks, power estimates always are the power can also go down with the scaling of the tech-
proportional to the Dynamic Range squared (equivalent nology, especially in matching dominated situations.
to 22ENOB, instead of 2ENOB) [5]. However, the overall Fact is that data converter performance has been dra-
power dissipation from ADCs built from those individ- matically improving over the past decade, especially in
ual blocks, seems to fit better to the Dynamic Range to the newer technologies. Many designs have taken
the power of one [2, 6]. This has to be because generally advantage of the higher speed, the higher density and
speaking adding one bit of resolution to the ADC the better matching that each step in the progress of
involves changing the architecture. That emphasizes the technology scaling brings us.
fact that the design landscape for ADCs is far from Generally speaking, as a result of the large num-
being smooth and is characterized by sharp transitions ber of transistors (approaching a billion) integrated in
going from one architecture to the next. large SoCs, typical "Analog" problems like matching,
In practice this often means that in order to find gate-leakage, subthreshold leakage and reduced intrinsic
the best suitable architecture, different designs have to gain, are becoming equally problematic for the digital
be worked out to a reasonable level of detail before an part of the chip. These problem will have to be and are
appropriate decision can be made. currently being solved, mostly in the technology
Special care should be taken when considering domain. Analog circuits will benefit from those solu-
published FOMs. In the last couple of years there has tions equally, or even more than digital circuits [5].
been a clear tendency to produce papers that can boast
of a very good FOM. For that reason often certain parts XI. CONCLUSIONS
of the power-dissipation are left out (like the SHA or Designing embedded data converter entails dif-
any form of input-buffer, the clock-drivers and the refer- ferent degrees of freedom compared to designing their
ences). However, what matters to the system is the total stand-alone counterparts: logic is for free, speed and
power-dissipation and a well optimized AFE design. matching are great, but on the other hand supply volt-
X. TECHNOLOGY SCALING [1-7] ages are low, process options are too costly and the
design has to yield at defect density. The constant fear
Many papers have been written about the effect that embedding analog circuits in large digital circuits
of technology scaling on Analog circuits [1 - 3, 5, 7] in will become impossible started two decades ago, and is
general and on Data Converters in particular [4, 6]. Up still alive but is not supported by evidence. Some of the
until about 0.8um technology, the supply voltage had fears have been put to rest (like cross-talk, which in the
been 5V for a long time. When it went down to 3.3V '90s was believed to kill the path to SoCs, is not even in
many people were afraid that would be the end of the the top 3 of current day embedded ADC designer prob-
dream of integrating analog with digital circuitry and lem-list anymore), but are being replaced by new fears
SoCs would never really happen. This discussion started (like low intrinsic gain and gate-leakage). But fact is
in the '90s and never really stopped. Every new drop in that many extremely large SoCs (more than 200mm 2

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
nd almost reaching a billion transistors) have been and IEEE International Solid-State Circuits Conference, vol.
XLII, pp. 316 - 317, February 1999. . .
still are in very large volume production (many millions [21] J. Ming and S. H. Lewis, "An 8b 80Msample/s pipelined
per month) and there is no sign this will stop in the near ADC with background calibration," IEEE International Solid-
State Circuits Conference, vol. XLIII, pp. 42 - 43, February
future. 2000.
The effect this has on the design of data convert- [22] B. Murmann and B. E. Boser, "A 12-bit 75-MS/s pipelined
ADC using open-loop residue amplification," IEEE Journal
ers is clear. More and more, "digital assisted" analog of Solid-State Circuits, vol. 38, pp. 2040 - 2050, December
design will become mainstream. The effect this has on 2003.
[23] E. Siragusa and 1. Galton, "A digitally enhanced 1.8-V 15-bit
the choice of data converter architecture is less obvious. 40-MSample/s CMOS pipelined ADC," IEEE Journal of
Although pipeline converters have become very popular Solid-State Circuits, vol. 39, pp. 2126 - 2138, December
2004.
in Nyquist-rate ADCs, almost all other architectures are [24] C. R. Grace, P. J. Hurst, and S. H. Lewis, "A 12b 80MS/s
still alive and it does not look like anyone of them will pipelined ADC with bootstrapped digital calibration," IEEE
International Solid-State Circuits Conference, vol. XLVII, pp.
win and obsolete the other architectures. Personal pref- 460 - 461, February 2004.
erences of and familiarity with certain architectures still [25] X. Wang, P. J. Hurst, and S. H. Lewis, "A 1~-bit 20-MsaI!1~lel
s pipelined analog-to-digital converter with nested digital
have a very strong influence on what choices designers background calibration," IEEE Journal of Solid-State Cir-
make. cuits, vol. 39, pp. 1799 - 1808, November 2004.
[26] M. Daito, H. Matsui, M. Ueda, and K. Iizuka, "A 14-bit 20-
MS/s pipelined ADC with digital distortion calibration,"
REFERENCES IEEE Journal ofSolid-State Circuits, vol. 41, pp. 2417 - 2423,
[1] K. Bult, "Analog Broadband Communication Circuits in Pure November 2006.
Digital Deep Sub-Micron CMOS", IEEE International Solid- [27] D. Draxelmayr, "A 6b 600MHz 10mW ADC array in digital
State Circuits Conference, vol. XLII, pp., February 1999. 90nm CMOS," IEEE International Solid-State Circuits Con-
[2] R.H. Walden, "Analog-to-Digital Converter Survey and Anal- ference, vol. XLVII, pp. 264 - 265, February 2004.
ysis", IEEE JSAC, vol. 17, No.4, 539-550, April 1999. [28] B. P. Ginsburg and A. P. Chandrakasan, "500-MS/s 5-bit ADC
[3] K. Bult, "Analog design in deep sub-micron CMOS", ESS- in 65-nm CMOS with split capacitor array DAC," IEEE Jour-
CIRC, pp., September 2000. nal ofSolid-State Circuits, vol. 42, pp. 739 - 747, April 2007.
[4] Y. Chiu and P.R. Gray, "Scaling of Analog-to-Digital Con- [29] S. M. Louwsma et al., "A 1.35 GS/s, 10 b, 175 mW time-
verters into Ultra-Deep Submircon CMOS", CICC, pp. 375- interleaved AD converter in 0.13 urn CMOS," IEEE Journal
382,2005. ofSolid-State Circuits, vol. 43, pp. 778 - 786? Apri~ 2008.
[5] K. Bult, "The Effect of Technology Scaling on Power Dissi- [30] B. P. Ginsburg and A. P. Chandrakasan, "HIghly Interleaved
pation in Analog Circuits", in Analog Circuit Design, M.Ste- 5-bit, 250-Msample/s, 1.2-m W ADC with redundant channels
yeart, A.H.M. v. Roermund and J.H. Huijsing, (eds), Springer, in 65-nm CMOS," IEEE Journal of Solid-State Circuits, vol.
2006. 43, pp. 2641 - 2650, December 2008.
[6] B. Murmann, "AID Converter Trends: Power Dissipation, [31] P. Schvan et al., "A 24GS/s 6b ADC in 90nm CMOS," IEEE
Scaling and Digitally Assisted Architectures", CICC, pp. 105- International Solid-State Circuits Conference, vol. 51, pp.
112,2008. 544 - 545, February 2008.
[7] http://www.itrs.net/links/2003ITRS/Home2003.htm . [32] Z. Cao, S. Van, and Y. Li, "A 32m W 1.25GS/s 6b 2b/step
[8] K.Uyttenhove et al., "Design Tec~niques an~ Implementation SAR ADC in Gl Sum CMOS," IEEE International Solid-
of an 8-bit 200-MS/s Interpolating/Averaging CMOS AID State Circuits Conference, vol. 51, pp. 542 - 543, February
Converter", IEEE Journal ofSolid-State Circuits" vol. 38, no. 2008.
3, pp. 483-494, March 2003. [33] R.J. v.d.Plassche and R.J.v.d.Grift, "A High-Speed 7 Bit AID
[9] S. Park et al., "A 3.5 GS/s 5-b Flash ADC in 90nm CMOS", Converter", IEEE Journal of Solid-State Circuits" Vol. SC-
IEEE CICC, pp. 489-492, 2006. 14, No.6, pp. 938-943, Dec. 1979.
[10] Y-Z. Lin et al., "A 5-bit 4.2-GS/s Flash ADC in 0.13-um [34] K. Bult and A. Buchwald, "An embedded 240-mW 10-b 50-
CMOS", IEEE CICC, pp. 213-216, 2007. MS/s CMOS ADC in l-mmo," IEEE Journal of Solid-State
[11] S.Park et al., "A 4-GS/s 4-bit Flash ADC in 0.18um CMOS", Circuits, vol. 32, pp. 1887 - 1895, December 1997.
IEEE Journal ofSolid-State Circuits" vol. 42, no. 9, pp. 1865- [35] M. Choe et al., "A 13-b 40-Msamples/s CMOS pipelined
1872, Sept. 2007. . folding ADC with background offset trimming," IEEE Jour-
[12] A. Dingwall and V. Zazzu, "An 8-MHz CMOS Subranging 8- nal of Solid-State Circuits, vol. 35, pp. 1781 - 1790, Decem-
bit AID Converter", IEEE Journal ofSolid-State Circuits" Vol. ber 2000.
SC-20, No.6, pp. 1138-1143, Dec. 1985. [36] G. Geelen and E. Paulus, "An 8b 600MS/s 200mW CMOS
[13] B. P. Brandt, J. Lutsky; "A 75-mW, 10-b, 20-MSPS CMOS folding AID converter using an amplifier preset technique,"
subranging ADC with 9.5 effective bits at Nyquist", IEEE IEEE International Solid-State Circuits Conference, vol.
Journal of Solid-State Circuits, vol. 34, pp. 1788 - 1795, XLVII, pp. 254 - 255, February 2004.
December 1999. [37] R. C. Taft et al., "A 1.8-V 1.6-GSample/s 8-b self-calibrating
[14] R. C. Taft, M. R. Tursi, "A 100-MS/s 8-b CMOS subranging folding ADC with 7.26 ENOB at Nyquist frequency," IEEE
ADC with sustained parametric performance from 3.8 V down Journal of Solid-State Circuits, vol. 39, pp. 2107 - 2115,
to 2.2 V", IEEE Journal of Solid-State Circuits, vol. 36, pp. December 2004.
331 - 338, March 2001. [38] R.C. Taft et al., "A 1.8V 1.0GS/s lOb Self-Calibrating Unified-Fo1d-
[15] J.Mulder et al., "A 21-mW 8-b 125-MSample/s ADC in 0.09- ing-Interpo1ating ADC with 9.1 ENOB at Nyquist Frequency", IEEE
mm2 0.13-um CMOS", IEEE Journal ofSolid-State Circuits" International Solid-State Circuits Conference, vol. LII, pp. 78
Vol. 39, No. 12, pp. 2116-2125, Dec. 2004. . - 79, February 2009.
[16] S. H. Lewis and P. R. Gray, "A pipelined 5-Msample/s 9-bIt [39] W. C. Black Jr. and D. A. Hodges, "Time interleaved con-
analog-to-digital converter", IEEE Journal ofSolid-State Cir- verter arrays," IEEE Journal of Solid-State Circuits, vol. 15,
cuits, vol. 22, pp. 954 - 961, December 1987. pp. 1022 - 1029, December 1980.
[[17] A. Karanicolas et al., "A 15-b I-Msample/s digitally self-cali- [40] Tat C. Choi, Robert W. Brodersen; Considerations for high-
brated pipeline ADC," IEEE Journal of Solid-State Circuits, frequency switched-capacitor ladder filters, IEEE Trans. Cir-
vol. 28, pp. 1207 - 1215, December 1993. cuit Syst., vol. CAS-27, pp. 545 - 552, June 1980. . .
[18] 1. E. Opris et al., "A single-ended 12-bit 20 Msample/s self- [41] B. Min et al., "A 69-mW 10-bit 80-MSample/s pipelined
calibrating pipeline AID converter," IEEE Journal of Solid- CMOS ADC," IEEE Journal of Solid-State Circuits, vol. 38,
State Circuits, vol. 33, pp. 1898 - 1903, December 1998. pp. 2031 - 2039, December 2003
[19] D. Fu, K. C. Dyer, S. H. Lewis, and P. J. Hurst, "A digital [42] K. Nagaraj et al, "A 250-mW, 8-b, 52-Msamples/s p~rallel­
background calibration technique for time-interleaved analog- pipelined AID converter with reduced number of amplifiers,"
to-digital converters," IEEE Journal of Solid-State Circuits, IEEE Journal of Solid-State Circuits, vol. 32, pp. 312 - 320,
vol. 33, pp. 1904 - 1911, December 1998. March 1997.
[20] O. E. Erdogan, P. J. Hurst, and S. H. Lewis, "A 12 b digital-
background-calibrated algorithmic ADC with -90dB THD,"

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.
[43] J. Li et al., "A lOb 170MS/s CMOS pipelined ADC featuring
84dB SFDR without calibration," Symp. VLSI Circuits Dig. ,
pp. 226 - 227, June 2006.
[44] E. Iroaga and B. Murmann, "A 12-bit 75-MS/s pipelined
ADC using incomplete settling," IEEE Journal ofSolid-State
Circuits, vol. 42, pp. 748 - 756, April 2007.
[45] L. Brooks and H. Lee, "A zero-crossing-based 8-bit 200 MS/s
pipelined ADC," IEEE Journal of Solid-State Circuits, vol.
42, pp. 2677 - 2687, December 2007.
[46] J. K. Fiorenza et al., "Comparator-based switched-capacitor
circuits for scaled CMOS technologies," IEEE Journal of
Solid-State Circuits, vol. 41, pp. 2658 - 2668, December 2006
[47] S. Shin et al., "A fully-differential zero-crossing-based 1.2V
lOb 26MS/s pipelined ADC in 65nm CMOS," Symp. VLSI
Circuits Dig. 22, pp. 218 - 219, June 2008.
[48] A. M. A. Ali et al., "A 14-bit 125 MS/s IF/RF sampling pipe-
lined ADC with 100 dB SFDR and 50 fsjitter," IEEE Journal
ofSolid-State Circuits, vol. 41, pp. 1846 - 1855, August 2006
[49] J. Li et al., "A 1.8-V 22-mW 10-bit 30-MS/s subsampling
pipelined CMOS ADC," 2006 IEEE Custom Integrated Cir-
cuits Conference, pp. 513 - 516, September 2006
[50] 1. Ahmed and D. A. Johns, "A high bandwidth power scalable
sub-sampling 10-bit pipelined ADC with embedded sample
and hold," Proceedings ofthe 33rd European Solid-State Cir-
cuits Conference, pp. 159 - 162, September 2007
[51] J. Li, X. Zeng, L. Xie, J. Chen, J. Zhang, and Y. Guo, "A 1.8-
V 22-mW 10-bit 30-MS/s pipelined CMOS ADC for low-
power subsampling applications," IEEE Journal of Solid-
State Circuits, vol. 43, pp. 321 - 329, February 2008
[52] 1. Ahmed and D. A. Johns, "A high bandwidth power scalable
sub-sampling 10-bit pipelined ADC with embedded sample
and hold," IEEE Journal of Solid-State Circuits, vol. 43, pp.
1638 - 1647, July 2008
[53] A.Panigada and 1.Galton, "A 130mW 100MS/s Pipelined
ADC with 69dB SNDR Enabled by Digital Harmonic Distor-
tion Correction", IEEE International Solid-State Circuits Con-
ference, vol. LII, pp. 162 - 163, February 2009.
[54] J. McNeill, M. Coln, and B. Larivee, "A split-ADC architec-
ture for deterministic digital background calibration of a 16b
1MS/s ADC," IEEE International Solid-State Circuits Con-
ference, vol. XLVIII, pp. 276 - 277, February 2005.
[55] J.Li et al., "A 0.9-V 12-mW 5-MSPS Algorithmic ADC With
77-dB SFDR", IEEE Journal ofSolid-State Circuits, vol. 40,
pp. 960 - 969, April 2005.
[56] J.McNeill et al., """Split ADC" Architecture for Determinis-
tic Digital Background Calibration ofa 16-bit I-MS/s ADC",
IEEE Journal ofSolid-State Circuits, vol. 40, pp. 2437 - 2445,
December 2005.
[57] 1.Ahmed and D.Johns, "An II-bit 45MS/s pipelined ADC
with rapid calibration of DAC errors in a multi-bit pipeline
stage", ESSCIRC 2007, pp. 147 - 150, September 2007.
[58] K.Poulton et al., "A 20GS/s 8b ADC with a 1MB Memory in
0.18um CMOS", IEEE International Solid-State Circuits
Conference, vol. XLVI, pp. 318-319, February 2003.
[59] S. Gupta, M. Choi, M. Inerfield, and J. Wang, "A 1GS/s lIb
time-interleaved ADC in 0.13urn CMOS," IEEE Interna-
tional Solid-State Circuits Conference, vol. XLIX, pp. 576 -
577, February 2006.
[60] C. Hsu, F. Huang, C. Shih, C. Huang, Y. Lin, C. Lee, and B.
Razavi, "An lIb 800MS/s time-interleaved ADC with digital
background calibration," IEEE International Solid-State Cir-
cuits Conference, vol. XL, pp. 464 - 465, February 2007.
[61] O. E. Agazzi et al., "A 90 nm CMOS DSP MLSD transceiver
with integrated AFE for electronic dispersion compensation
of multimode optical fibers at 10 Gb/s," IEEE Journal of
Solid-State Circuits, vol. 43, pp. 2939 - 2957, December 2008

Authorized licensed use limited to: JAWAHARLAL NEHRU TECHNOLOGICAL UNIV. Downloaded on September 29,2022 at 10:16:52 UTC from IEEE Xplore. Restrictions apply.

You might also like