Anda di halaman 1dari 12


12, DECEMBER 2001


A 6-b 1.3-Gsample/s A/D Converter in

0.35-m CMOS
Michael Choi and Asad A. Abidi, Fellow, IEEE

AbstractA 6-b Nyquist A/D converter (ADC) that converts

at 1.3 GHz is reported. Using array averaging and a wideband
track-and-hold, a 6-b flash ADC achieves better than 5.5 effective
bits for input frequencies up to 630 MHz at 1 Gsample/s, and five
effective bits for 650-MHz input at 1.3 Gsample/s. Peak INL and
DNL are less than 0.35 LSB and 0.2 LSB, respectively. This ADC
consumes about 500 mW from 3.3 V at 1 Gsample/s. The chip occupies 0.8-mm2 active area, fabricated in 0.35- m CMOS.
Index TermsA/D converter, CMOS analog integrated circuits,
comparators, flash converter, offset averaging, read channel, resistor averaging network, spatial filter, track-and-hold.


IGITAL receivers for high-bit-rate communications

are spurring on the conversion rate of analog-to-digital
converters (ADCs). State-of-the-art disk-drive read channels
and high-speed Ethernet signals use partial-response signaling,
which requires 6-b resolution at conversion rates of 1 GHz
and beyond [1]. Furthermore, continuous DVD playback and
certain Ethernet uses permit no idle times when the linearity of
the ADC may be self-calibrated by autozeroing [2].
Fig. 1 shows the block diagram of an ADC embedded in the
front end of a 1-Gbit/s PRML read channel. In this architecture,
the amplified read waveform is sampled and digitized for detection. Accurate timing and gain control requires a dynamic range
higher than 45 dB. The analog delay line can degrade the signal
dynamic range by 10 dB. Therefore, the dynamic range at the
sampler output should be higher than 55 dB which defines the
ADCs track-and-hold (T/H) linearity requirement. Converter
latency is not important here, because the ADC is not part of
the timing and gain control loops. For PRML waveforms, the
6-b ADC should convert waveforms occupying half the Nyquist
bandwidth with 5.5 effective-bit linearity.
The literature describes the flash architecture and variations
used in many high-speed low-resolution ADCs. The fastest reported 6-b ADC operates at 4 Gsample/s, using a folding architecture with on-chip T/H and implemented in a GaAs HBT technology [3]. Folding architectures lead to low latency and power,
but the requisite analog preprocessing of input signal degrades
dynamic performance [4]. If a T/H is not present, autozeroing
needs series-input capacitors that limit resolution bandwidth,
and also requires idle time for offset cancellation [5], [6].
Background autozeroing enables continuous conversion, but reManuscript received April 10, 2001; revised June 19, 2001.
The authors are with the Department of Electrical Engineering, University of
California, Los Angeles, CA 90095-1594 USA (e-mail:
Publisher Item Identifier S 0018-9200(01)09318-0.

Fig. 1. 1-Gsample/s 6-b A/D converter intended for use in a disk-drive read

quires complex clocking that is incompatible with high-speed

operation [7].
An input T/H improves the resolution bandwidth if wideband signals are to be digitized with the specified resolution
[3], [8]. Instead of acquiring samples of the input waveform,
a distributed T/H array [9] captures the zero crossings at the
output of a flash-like array of preamplifiers. The T/H circuits in
this arrangement need not be linear, which simplifies the circuit,
but like any parallel scheme, they suffer dynamic inaccuracies
due to skews in the clock signals distributed across the array. At
high input frequencies, a single front-end T/H with sufficient
linearity will outperform the distributed T/H [9], [10].
This paper reports on a 6-b ADC without autozero or self-calibration, which digitizes a 630-MHz input tone with a linearity
of 5.5 effective bits at 1 Gsample/s, and a 650-MHz input with
five effective bits at 1.3 Gsample/s. Enabling this performance
for a CMOS ADC are a simple yet accurate T/H circuit and
analog techniques to improve accuracy at high conversion rates
and wide input-signal bandwidths. Section II presents the architecture and key ideas to improve the sampling rate and resolution bandwidth. Section III describes resistive averaging to improve linearity and speed. Circuit implementation of the major
building blocks is described in Section IV. Finally, experimental
results are discussed in Section V.
As Fig. 2 shows, this is a flash ADC. Balanced circuits in
the T/H, reference ladder, and other preamplifiers largely cancel
second-order distortion. The T/H is designed for 8-b linearity,
but with the highest possible bandwidth. The preamplifier array
with gain of 3 senses the difference between the differential
input signal and the differential threshold to drive the latched

00189200/01$10.00 2001 IEEE



cient dummy preamplifiers maintains translational symmetry of

the array across the input full scale.
B. Speed Improvement With Averaging Network
It is by now well known that averaging improves accuracy.
What is not as well appreciated, as explained below, is that averaging also improves speed.
The input random offset voltage is inversely proportional to
the square root of the transistor gate area [15]. Therefore, preamplifier FET size may be related to the target ADC resolution as

Fig. 2. Block diagram of 6-b flash A/D converter.

comparator. However, the preamplifiers in the array suffer from
random offsets. Collective averaging of the preamplifier outputs
across a properly designed resistor network lowers the impact
of the offsets, improving accuracy of the threshold comparison
without degrading bandwidth [13]. A comparator must quickly
recover from large overdrive when the changing input voltage
rapidly approaches the comparators threshold from far away.
To aid this, the gain of the comparators input differential pair
is about one, which widens its bandwidth. The resulting net amplification of 3 is not sufficient to overcome the dynamic random
offsets in the regenerative latch for 6-b accuracy. Resistor averaging is also used within the comparators to lower the impact
of offsets. Interpolation is not used in this work, because it degrades bandwidth [12], [13]. The second comparator array provides rail-to-rail logic swing for the digital back end consisting
of the SR latch and ROM-based digital encoder.

References [13], [14] show that the optimum averaging network

lowers random offset by up to 3 , with a spatial impulse re18 nodes wide. With this network connected to
the outputs of a preamplifier array, a given accuracy is maintained with 9 smaller FETs than if no averaging were used.
The bandwidth as set by the output pole of this differential pair,
shown in Fig. 4, is
With averaging, the wiring and junction capacitance are 9
lower due to transistor scaling (3 less
). For the same bias
must be 3 larger for equal
current, the load resistance
voltage gain, which means that



A. Accuracy Improvement With Averaging Network

Fig. 3 shows a preamplifier array with averaging resistors
connecting adjacent output nodes. In this work, offset averaging
is treated and optimized based on the concept of spatial filtering

network forms a spatial filter with im[13], [14]. The

pulse response

is the parasitic capacitance of the averaging netwhere

work. Comparing the two terms from (3) and (4)

where represents the index of differential pairs in the array.
determines the width of the impulse response.
The ratio
An array of differential pairs injects two different types of
stimuli currents in the averaging network. First, differential
limited by the linear region of the
signal currents
differential pair enter the network. Second, noise currents
due to transistor random mismatches also enter the
averaging network. A well-designed averaging network should
filter out the random currents, without losing valuable signal
current, nor should the network lower bandwidth. Averaging is
(the number of differential pairs in the
optimum when
unclipped linear region of their characteristic) is greater than
the impulse response width of the network
Dummy preamplifiers are inserted to compare the input signals with thresholds beyond the actual full scale, so that the tails
of the averaging networks impulse response remain undistorted
at the ends of the full scale. A circular arrangement with suffi-

The second term in (5) is much less than 1, while the third term
is close to 1. Therefore, use of the averaging network raises
preamplifier bandwidth by 3 .
A. Track-and-Hold
An input T/H improves the dynamic performance of an
ADC. By holding the analog sample static during digitization,
the T/H largely removes errors due to skews in clock delivery to
a large number of comparators, limited input bandwidth prior
to latch regeneration, signal-dependent dynamic nonlinearity,
and aperture jitter [11]. The on-chip T/H circuit shown in Fig. 5
precedes the flash quantizer. This simple circuit, consisting
of a passive nMOS switch connected to a sampling capacitor
through a dummy switch [16], [17], offers just the required
linearity. The dummy switch lowers the common-mode jump
after the track-to-hold transition. The main sources of distortion


Fig. 3.


Preamplifier array with resistor averaging network.

effectively cancel charge feedthrough, even with process variations.

Sample jitter due to the input-dependent switch opening is another distortion source and is described by the following output
equations with a differential sine input.

Fig. 4.


Single differential amplifier.

Fig. 5. On-chip track-and-hold.

are signal-dependent charge injection on switch opening,

nonlinearity of the source follower, and dynamic current into
the signal-dependent input capacitance of the quantizer.
FET charge injection and clock feedthrough cause distortion by adding/removing charge on the hold capacitor when
the switch disconnects the signal source [18]. A dummy switch
driven by the complement of the switch clock lowers these effects [19]. Because both the source and drain of the dummy
ratio is initially
switch are connected to the hold node, its
chosen as half the size of the switch, and then tuned through simulation. A low input common-mode voltage, 0.5 V, enables use
of only nMOSFETs for the sampling and dummy switches. Unlike a complementary switch, nMOS (or pMOS)-only switches

is the maximum clock voltage,

is the transiwhere
tion time of a clock edge, is the input frequency, and is
(3.3 V),
(0.15 ns),
the input amplitude [20]. For given
(500 MHz), and (0.4 V), the total harmonic distortion is
68 dB for the differential output, and 25 dB for the singleended.
The low-input common-mode voltage of 0.5 V allows a larger
gate overdrive to turn on the switch, which lowers track-mode
distortion due to nonlinear channel resistance. The source is tied
to the well in the pMOS source follower to eliminate nonlinear
body effect. Simulations show that when acquiring samples of a
250-MHz full-scale sine wave at 1 Gsample/s, the T/H delivers
samples to the quantizer input with third-harmonic distortion of
about 60 dBc.
B. Preamplifier
The preamplifier stage should be wideband and provide sufficient gain to overcome comparator offsets. It should also recover
from large overdrive within one clock cycle. An open-loop lowgain-stage amplifier is gain-bandwidth limited, therefore unsuitable for use at high speeds because of poor overdrive recovery
[6], [21][23].
The following analysis addresses the fundamental limitation
of an open-loop single-pole amplifier stage in overdrive recovery and justifies quantitatively that an amplifier with a reset
switch meets the speed requirement. Consider a simple onestage single-pole amplifier, shown in Fig. 4. The preamplifier is
. With a step input applied to
completely unbalanced at



Fig. 8. Required GBW of a single-pole amplifier to amplify by 3 within one

period of 1.3 GHz.

Fig. 6. Output transient response to a step input.

Fig. 7.

Required amplifier bandwidth given the overdrive recovery time.

the preamplifier stage, the output transient is shown in Fig. 6.

The step response of the amplifier is given by
is the voltage difference bewhere is the voltage gain,
tween the input and reference tap, is the tail current, and is
the time constant. The overdrive recovery time is calculated by
of (7) to zero.

Fig. 9. Achievable GBW of an NFET versus effective gate voltage

0.35-m CMOS.


overdrive recovery time for

to change polarity.
1 GHz is required to recover from overdrive in 0.35 ns.
Now assume that the output step response does not settle in
one sampling period as shown in Fig. 6. The gain-bandwidth
requirement to obtain the desired gain (
is given by
can be derived from (7). The output voltage at

Solving (8) for

results in

Solving (11) for


In order to determine the relationship between
with fixed bias current and transistor size, substitute
into (9), where
Solving this gives

The gain-bandwidth product (GBW) required to obtain gain of

) within one period of 1.3-GHz sampling rate is
, multiplying (12) by the dc gain
shown in Fig. 8 for various
. GBW is optimum when the dc gain is about 4. Given
of 0.1 V, GBW should be higher than 2.5 GHz. However, the
GBW is limited by the FET channel length as follows:

mV and
V, Fig. 7
shows the amplifier stage bandwidth requirement versus the

is a constant that defines

. Fig. 9 shows the
based on extracted simulaachievable GBW with given



voltage gain lowers the output common-mode which can cause

the input transistors to enter the triode region. For preamplification by 3 within a half period of 1.3 GHz, a dc voltage gain of
4 is selected, which needs 2.2-GHz GBW. As shown in Fig. 9,
this can be obtained with
C. First-Stage Comparator

Fig. 10.

Preamplifier with a reset switch.

Fig. 11.

Required GBW to obtain gains of 3, 4, and 5 with a reset amplifier.

tions. Therefore, the GBW specified in Fig. 8 cannot be met

with 0.35- m CMOS.
However, it is possible to surmount this apparently fundamental limit by exploiting the clocked nature of this circuit.
is inserted between the two output nodes,
A reset switch
as shown in Fig. 10. While the T/H is in track mode, the reset
switch is turned ON to erase the residual voltage from the previous sample. When the T/H is in hold mode, the reset switch is
turned OFF. Similar to the previous analysis, the step response
for the amplifier with the reset switch is given by
The desired voltage gain should be made available at
because the amplifier is in reset mode for half the period. The
is given by
output voltage at
Solving (15) for


The GBW required to obtain amplification of 3, 4, and 5 at

of 1.3-GHz sampling rate is shown
the end of a half period
in Fig. 11. The higher the dc gain, the lower the required GBW.
However, the offset averaging network needs a well-controlled
load resistor. As well, an increase in load resistor for higher

Fig. 12 shows the first-stage comparator. This circuit consists

and a latch pair

of an input differential pair

. Unlike
both sharing the diode-connected MOS load
a MOS triode region resistor, this load largely decouples the
signal gain and output common-mode voltage. Due to the circular and symmetric nature of the network, nonlinearity of the
diode loads does not offset zero-crossing at the output of the resistive averaging network.
is low and the differential output folIn reset mode,
lows the differential input. While the comparators input is being
lowers the first comsampled to its output, a shorting switch
parator voltage gain to less than one and erases memory of the
previous decision. A similar comparator has been reported previously [24], [25], which applies a short pulse to the reset switch
to improve voltage gain. However, it is impractical to generate
a short pulse at 1.3 GHz, when the clock transitions alone consume 40% of the period.
to turn on the
This comparator uses the standard
. Fixed gain-bandwidth product implies that
shorting switch
the low voltage gain during the comparators reset mode gives a
high output bandwidth for fast erase. However, the gain should
be large to overcome dynamic offset in the regenerative latch
[6], [26][29] of as much as 40-mV rms. Rather than raise the
voltage gain and degrade speed, we have discovered that resistor averaging across an array of regenerating latches lowers
dynamic offsets as well. Use of averaging in the latch stage increases regeneration time constant that necessitates cascaded
comparators to lower bit-error rate. However, the overall power
reduction and sampling rate increase due to averaging surpasses
this minor drawback as discussed in the following section.
D. Two Stages of Averaging
Fig. 13 shows the complete block diagram of the quantizer
with cascaded averaging. This flash ADC contains an array of
63 preamplifiers with nine dummies at each end of the array
for proper averaging at the extremes of the input full scale. The
differential averaging network consists of diode pMOS loads of
the preamplifiers and polysilicon resistors connected between
neighbors in the array. The averaging network in the comparator
array is similar.
Monte Carlo simulations verify the benefits of resistive averaging. The transistor threshold voltage and transconductance are
randomly distributed according to the sigma values provided by
the fabricator and modeled by (17) and (18). The Monte Carlo
transient simulations are performed on the two-stage averaging
resistor network in Fig. 13.



Fig. 12.


Circuit of first-stage comparator.

Fig. 14. Dynamic Monte-Carlo simulation of input-referred random offset of

latched comparator.

Fig. 13. Two stages of averaging resistor networks.

In the ideal case with no transistor mismatch, the comparator
output changes polarity when a slow ramp input crosses the
corresponding threshold. With random variations added to the
transistors and 10% mismatch in the load capacitance, Fig. 14
shows that the time the comparator changes polarity is randomly
dispersed with respect to when the input crosses zero. For each
Monte Carlo simulation, the input voltage is measured when the
output changes its polarity. The sigma of input-referred random
offset is obtained by curve-fitting the corresponding input voltages to the Gaussian distribution. Fig. 14 is a composite plot of
60 Monte Carlo transient analyses.
Without averaging, the input-referred random offset in one
comparator chain is about 11 mV, as shown in Fig. 15(a). With

offset averaging applied only to the preamplifier stage, shown

in Fig. 15(b), the offset sigma is reduced to 9 mV. This is not
sufficient for 6-b resolution. The dominant offset source is the
dynamic offset in the first latch. With offset averaging applied
both to the preamplifier and comparator stages, as shown in
Fig. 15(c), the input-referred random offset sigma value is lowered to 3.7 mV, which is sufficient for 6-b linearity. This sigma
value can be derived analytically as follows.


with simulated preamplifier random offset (
mV), both averaging factors
latch random offset (
) of 3, and preamplifier gain (
We conclude that averaging after the preamplifier array and
dynamic comparator array reduces the overall input-referred
random offset by about 3 .



configured from cross-coupled CMOS inverters (

steer the tail current from one side to the other, speeding up
to couple the nMOSFETs, one
regeneration. Without
to below
output node needs to drop its voltage from
to turn off the corresponding nMOS transistor. The SR latch is
the same structure as the second-stage comparator without the
tail current source.
F. Digital Encoder



Fig. 15. Results of Monte Carlo simulation. (a) Without any averaging.
(b) With averaging in the preamplifier stage only. (c) With averaging applied
both to the preamplifier and comparator stages.

Suppose, hypothetically, that averaging is only applied to

the preamplifier array, and the preamplifier gain
raised to overcome latch dynamic offsets. This gain must be

This has important repercussions on the power consumption.
To raise the gain 3 in a resistor-loaded single-stage amplifier
whose output pole frequency, set by the junction capacitance, is
must be raised 3 . This implies
to be kept fixed, the FET
9 higher current. Therefore, cascaded averaging saves almost
an order of magnitude in preamplifier power.
E. Second-Stage Comparator
Fig. 16 shows the second-stage comparator. A single-phase
clock is applied from this stage throughout the digital back
end; multiple phases with the need to accommodate skews
would limit the highest clock frequency [30]. The second stage
provides rail-to-rail output to the SR latch. As in the case of the
preamplifier and the first-stage comparator, overdrive recovery
time also limits highest ADC clock frequency. During reset
mode, its output is reset through two parallel discharge paths
for fast overdrive recovery. In the next half clock cycle of
, and
regeneration, differential pair nMOSFETs (

Many digital encoding schemes have been developed to suppress glitch errors caused by the comparator metastability [31],
[32] and bubble errors in the thermometer code [7], [34][38].
Metastability errors occur when nonbinary comparator levels
drive the digital encoder and produce senseless outputs. Bubble
errors result from three major sources. First, the overall input-referred random offset greater than 0.5 LSB can switch the order
of the two adjacent thresholds and create a bubble error. Second,
for the comparators without a single front-end T/H, slew-dependent sampling and clock dispersion [39] increase bubble-error
probability. Finally, the propagation-delay variations through
the preamplifier stage as a result of limited bandwidth and fast
input frequency [7], [25] worsen the bit-error rate due to the
bubble errors.
In this work, the cascaded regeneration times of two ranks of
comparators and SR latch guarantee a metastable error rate of
at 1 Gsample/s. The front-end T/H eliminates
less than 10
the clock skew-dependent bubble errors. The reset switches in
the preamplifier and the first-stage comparators suppress the
bubble errors caused by inadequate overdrive recovery.
With the three major bubble error sources removed, a ROMbased encoder maps thermometer code output from the flash
array into quasi-Gray code (Fig. 17) [40], [41]. This arrangement requires only two-input NAND gates, yet suppresses bubble
errors equivalently to a conventional Gray encoder, which uses
slower three-input NAND gates. To operate at 1-GHz sampling
rate and beyond, the D-flip-flops are implemented with true
single-phase clocked (TSPC) circuits [42], [43]. The encoded
output is decimated by 16 on chip to enable easy acquisition
by a logic analyzer. The output buffers are differential circuits
driving 100- loads.
G. Clock Generator
In this ADC, clock jitter randomly modulates the periodic
sampling instants of the T/H. Nonuniform sampling raises the
noise floor of the digitized system, degrading signal-to-noise
ratio (SNR). Clock jitter is increasingly a concern relative to
the short period of high-speed ADCs. Given a sinusoidal input
waveform with amplitude and radian frequency , SNR due
to clock jitter only is
is the rms clock jitter [44]. According to (21), to obwhere
tain SNR of 38 dB (the ideal SNR for 6-b quantization of a
sinewave) at input frequency of 650 MHz, the rms clock jitter
should be less than 3 ps.



Fig. 16.

Circuit of second-stage comparator.

Fig. 17.

High-speed digital encoder.

reflections which may upset duty cycle, into

[14]. The size of each CMOS inverter buffer chain is based on
and 1.5 pF for
. When
the clock load: 5.9 pF for
this ADCs measurement (35-dB SNDR with 650-MHz input in
Fig. 23) is applied to (21), the deduced clock jitter is less than
4.3 ps.

Fig. 18.

Clock generation circuit.

Low-noise methods taken from analog circuit design are applied to the clock generator (Fig. 18). The circuits convert a differential sinewave input, terminated in 50 to suppress off-chip

The chip occupies 2.2

2.2 mm with 0.8-mm active area
fabricated through ST Microelectronics using MOSFETs only
in the 0.35- m four-metal single-poly BiCMOS6M technology.
The process offers double-metal linear capacitors and precise
poly resistors. More than half of the total pads (33 out of 64) are
assigned to power and ground to lower series inductance. Digital
and analog supplies are separately connected to on-chip decoupling capacitors (1.2 nF for digital power and 0.4 nF for analog
power). Extracted simulation shows that total 600-pF decoupling capacitor with 5.6-nH inductance causes 100 mV power
and ground bounce. The larger on-chip decoupling capacitance



Fig. 19.

Microphotograph of ADC active area, and connection topology of preamplifiers.

Fig. 20.

ADC test setup.

avoids unexpected performance degradation due to power and

ground bounce.
Fig. 19 shows a microphotograph of the ADC layout. The
averaging network connects the two edge preamplifiers in the
array. The folded reference ladder provides differential reference voltages to the preamplifier array so that the two preamplifiers at the edges are positioned next to each other. This equalizes the parasitic capacitance at each node in the array. The first
comparator array with averaging network is laid out in the same
way. The digital back end, pipelined for high-speed operation,
occupies about 50% of the active chip area.
The entire chip was extracted and simulated using mixedsignal techniques, and the reported performance is from the
first silicon. Fig. 20 shows the test setup of the ADC. The
die is mounted in a 64-pin TQFP for all tests. To measure
dynamic performance accurately, the ADC is clocked with a

Fig. 21.

Measured INL and DNL at 1 Gsample/s.

low-phase-noise sinewave synthesizer, appropriately bandpass

filtered to reduce jitter and harmonics. Passive phase splitters



Fig. 22.

Fig. 23.


Reconstructed spectra. (a) 630-MHz at 1 Gsample/s. (b) 650-MHz at 1.3 Gsample/s.


Measured dynamic performance. (a) SNDR versus input frequency. (b) SNDR and SFDR versus conversion rate with 100-MHz input.

generate differential analog input and clock signals in response

to the single-ended versions emanating from the frequency
synthesizers. Bias-Ts are used for ac coupling. The signals
are properly terminated to on-chip 50- poly resistors. A logic
analyzer captures the 16 decimated ADC output, and the data
is transferred to a PC for analysis.
Integral nonlinearity (INL) and differential nonlinearity
(DNL) are calculated from histograms of ADC output data.
When digitizing a 630-MHz sinusoid at 1 Gsample/s, peak
DNL is less than 0.2 LSB, and peak INL is below 0.35 LSB
(Fig. 21). Fig. 22 shows the reconstructed spectra. With a
630-MHz input at 1 Gsample/s, the ADC is 5.5 effective-bit
accurate. With 650-MHz input converting at 1.3 Gsample/s,
the ADC is five effective-bit accurate. Both of these figures
are consistent with 6-b linearity and indicate that the cascaded
resistor averaging effectively suppresses threshold errors to
efficiently attain 6-b linearity.
The ADCs dynamic performance shows an effective resolution bandwidth exceeding Nyquist input frequency. The
signal-to-noise-plus-distortion ratios (SNDRs) at 1.025 and

1.312 GHz conversion rate with various input frequencies are

plotted in Fig. 23(a). The reconstructed SNDR is a flat 35 dB
up to 630-MHz input for 1-GHz sampling rate, and 32 dB up
to 650-MHz input for 1.3-GHz sampling rate. Fig. 23(b) shows
SNDR with fixed low input frequency (100 MHz) but swept
conversion rate. Therefore, SNDR is above 5.5 effective bits up
to a maximum conversion rate of 1.34 GHz.
Bit-error rate is measured by digitizing a full-scale low-frequency sinewave at 1 Gsample/s which changes by at most
1 LSB in successive samples. Ten thousand iterations of one
million output data are captured by a logic analyzer. A bit error
is detected when the digitized output shows a change of more
than 2 LSB in successive samples. No error was detected in
this test to meet the bit-error rate requirement ( 10 ) of
partial-response maximum-likelihood (PRML) read channel.
Excluding output buffers, this ADC consumes about 500 mW
from 3.3 V at 1 Gsample/s, and 10% more at 1.3 Gsample/s. The
logic circuits and clock buffers consume half the power, and this
portion of the total power will scale down with technology. The
performance summary is given in Table I.


Fig. 24.



Comparisons for (a) reported 5 effective number of bits (ENOB) limits of 6-b CMOS ADCs and (b) reported 5.5 ENOB limits of 6-b CMOS ADCs.


Fig. 24 compares the performance of this ADC with other
CMOS 6-b converters reported in recent years. It is apparent
how the ADC reported here advances the state of the art. The
key features of this work are now summarized. A T/H for this
6-b ADC enables beyond-Nyquist input up to 1.3-GHz conversion rate. Reset switches in the preamplifier and comparator arrays give fast overdrive recovery. The second comparator is the
fastest such CMOS circuit with rail-to-rail output. In this work,
averaging networks are used in three places: first, to lower the
impact of preamplifier static offset; second, to lower the latch
dynamic offset; and third, to improve the preamplifier bandwidth by 3 . These ideas have led to a compact CMOS ADC
with unprecedented dynamic performance.
[1] S. Altekar et al., A 700-Mb/s BiCMOS read channel integrated circuit,
in IEEE Int. Solid-State Circuits Conf., Feb. 2001, pp. 184185.
[2] S. Gotoh et al., A mixed-signal 0.18-m CMOS SOC for DVD systems
with 432-MS/s PRML read channel and 16-Mb embedded DRAM, in
IEEE Int. Solid-State Circuits Conf., Feb. 2001, pp. 182183.

[3] K. Poulton et al., A 6-b 4-GSa/s GaAs HBT ADC, IEEE J. Solid-State
Circuits, vol. 30, pp. 11091118, Oct. 1995.
[4] M. P. Flynn and B. Sheahan, A 400-MSample/s 6-b CMOS folding and
interpolating ADC, in IEEE Int. Solid-State Circuits Conf., Feb. 1998,
pp. 150151.
[5] K. Nagaraj et al., A 700-MSample/s 6-b read channel A/D converter
with 7-b servo mode, in IEEE Int. Solid-State Circuits Conf., Feb. 2000,
pp. 426427.
[6] I. Mehr and D. Dalton, A 500-MSample/s, 6-b Nyquist-rate ADC for
disk-drive read-channel application, IEEE J. Solid-State Circuits, vol.
34, pp. 912920, July 1999.
[7] S. Tsukamoto, W. Schofield, and T. Endo, A CMOS 6-b,
400-MSample/s ADC with error correction, IEEE J. Solid-State
Circuits, vol. 33, pp. 19391947, Dec. 1998.
[8] Y. Tamba and K. Yamakido, A CMOS 6-b 500-MSample/s ADC for a
hard disk-drive read channel, in IEEE Int. Solid-State Circuits Conf.,
Feb. 1999, pp. 324325.
[9] G. Geelen, A 6-b 1.1-GSample/s CMOS A/D converter, in IEEE Int.
Solid-State Circuits Conf., Feb. 2001, pp. 128129.
[10] M. Choi and A. A. Abidi, A 6-b 1.3-GSample/s A/D converter in
0.35-m CMOS, in IEEE Int. Solid-State Circuits Conf., Feb. 2001,
pp. 126127.
[11] R. J. van de Plassche, Integrated Analog-to-Digital and Digital-toAnalog Converters. Norwood, MA: Kluwer, 1994.
[12] K. Sushihara et al., A 6-b 800-MSample/s CMOS A/D converter, in
IEEE Int. Solid-State Circuits Conf., Feb. 2000, pp. 428429.
[13] H. Pan, M. Segami, M. Choi, J. Cao, and A. A. Abidi, A 3.3-V 12-b
50-MS/s A/D converter in 0.6-m CMOS with over 80-dB SFDR,
IEEE J. Solid-State Circuits, vol. 35, pp. 17691780, Dec. 2000.
[14] H. Pan, A 3.3-V 12-b 50-MS/s A/D converter in 0.6-m CMOS with
over 80-dB SFDR, Ph.D. dissertation, Univ. of California, Los Angeles,
CA, Dec. 1999.
[15] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, Matching
properties of MOS transistors, IEEE J. Solid-State Circuits, vol. 24, pp.
14331440, Oct. 1989.
[16] D. Wei, D. Sun, and A. Abidi, A 300-MHz mixed-signal FDTS/DFE
disk read channel in 0.6-m CMOS, in IEEE Int. Solid-State Circuits
Conf., Feb. 2001, pp. 186187.
[17] P. Pai, A. Brewster, and A. Abidi, A 160-MHz analog front-end IC for
EPR-IV PRML magnetic storage read channels, IEEE J. Solid-State
Circuits, vol. 31, pp. 18031816, Nov. 1996.
[18] G. Wegmann, E. Vittoz, and F. Rahali, Charge injection in analog MOS
switches, IEEE J. Solid-State Circuits, vol. SC-22, pp. 10911097, Dec.
[19] C. Eichenberger and W. Guggenbuhl, Dummy transistor compensation
of analog MOS switches, IEEE J. Solid-State Circuits, vol. SC-24, pp.
11431146, Aug. 1989.
[20] P. Lim and B. Wooley, A high-speed sample-and-hold technique using
a Miller hold capacitance, IEEE J. Solid-State Circuits, vol. 26, pp.
643651, Apr. 1991.


[21] K. Uyttenhove, A. Marques, and M. Steyaert, A 6-b 1-GHz acquisition

speed CMOS flash ADC with digital error correction, in IEEE Custom
Integrated Circuits Conf., May 2000, pp. 249252.
[22] Y. Tamba and K. Yamakido, A CMOS 6-b 500-MSample/s ADC for a
hard disk-drive read channel, in IEEE Int. Solid-State Circuits Conf.,
Feb. 1999, pp. 324325.
[23] K. Sushihara et al., A 6-b 800-MSample/s CMOS A/D converter, in
IEEE Int. Solid-State Circuits Conf., Feb. 2000, pp. 428429.
[24] I. Mehr and D. Dalton, A 500-MSample/s, 6-b Nyquist-rate ADC for
disk-drive read-channel application, IEEE J. Solid-State Circuits, vol.
34, pp. 912920, July 1999.
[25] D. Dalton et al., A 200-MSPS 6-b flash ADC in 0.6-m CMOS, IEEE
Trans. Circuits Syst., vol. 45, pp. 14331444, Nov. 1998.
[26] A. Yukawa, A CMOS 8-b high-speed A/D converter IC, IEEE J. SolidState Circuits, vol. SC-20, pp. 775779, June 1985.
[27] B. S. Song, S. H. Lee, and M. F. Tompsett, A 10-b 15-MHz CMOS
recycling two-step A/D converter, IEEE J. Solid-State Circuits, vol. 25,
pp. 13281338, Dec. 1990.
[28] P. Setty, J. Barner, J. Plany, H. Burger, and J. Sonntag, A 5.75-b 350MSample/s or 6.75-b 150-MSample/s reconfigurable flash ADC for a
PRML read channel, in IEEE Int. Solid-State Circuits Conf., Feb. 1998,
pp. 148149.
[29] K. Bult and A. Buchwald, An embedded 240-mW 10-b 50-MS/s
CMOS ADC in 1-mm , IEEE J. Solid-State Circuits, vol. 32, pp.
18871895, Dec. 1997.
[30] J. M. Rabaey, Digital Integrated Circuits. Englewood Cliffs, NJ: Prentice Hall, 1996.
[31] K. Ono, T. Matsuura, E. Imaizumi, H. Okazawa, and R. Shimokawa,
Error suppressing encode logic of FCDL in a 6-b flash A/D converter,
IEEE J. Solid-State Circuits, vol. 32, pp. 14601464, Sept. 1997.
[32] C. Portmann and T. Meng, Power-efficient metastability error reduction
in CMOS flash A/D converters, IEEE J. Solid-State Circuits, vol. 31,
pp. 11321140, Aug. 1996.
[33] M. Steyaert, R. Roovers, and J. Craninckx, A 100-MHz 8-b CMOS
interpolating A/D converter, in IEEE Custom Integrated Circuits Conf.,
May 1993, pp.
[34] Y. Gendai, Y. Komatsu, S. Hirase, and M. Kawata, An 8-b 500-MHz
ADC, in IEEE Int. Solid-State Circuits Conf., Feb. 1991, pp. 172173.
[35] A. Matsuzawa, S. Nakashima, I. Hidaka, S. Sawada, H. Kodaka, and
S. Shimada, A 6-b 1-GHz dual-parallel A/D converter, in IEEE Int.
Solid-State Circuits Conf., Feb. 1991, pp. 174175.
[36] A. Matsuzawa, Y. Kitagawa, I. Hidaka, S. Sawada, M. Kagawa, and M.
Kanoh, An 8-b 600-MHz flash A/D converter with multistage duplex
gray coding, in Symp. VLSI Circuits, May 1991, pp. 113114.
[37] C. W. Mangelsdorf, A 400-MHz input flash converter with error correction, IEEE J. Solid-State Circuits, vol. 25, pp. 184190, Feb. 1990.
[38] V. Garuts, Y. Yu, E. Traa, and T. Yamaguchi, A dual 4-b 2-GS/s full
Nyquist analog-to-digital converter using a 70-ps silicon bipolar technology with borosenic-poly process and coupling-base implant, IEEE
J. Solid-State Circuits, vol. 24, pp. 216222, Apr. 1989.
[39] B. Razavi, Principles of Data Conversion System Design. New York:
IEEE Press, 1995.
[40] Y. Akazawa, A. Iwata, T. Wakimoto, T. Kamato, H. Nakamura, and H.
Ikawa, A 400-MSPS 8-b flash AD conversion LSI, in IEEE Int. SolidState Circuits Conf., Feb. 1987, pp. 9899.


[41] T. Wakimoto, Y. Akazawa, and S. Konaka, Si bipolar 2GS/s 6b flash

A/D conversion LSI, in IEEE Int. Solid-State Circuits Conf., Feb. 1988,
pp. 232233.
[42] J. Yuan and C. Svensson, New single-clock CMOS latches and flipflops with improved speed and power savings, IEEE J. Solid-State Circuits, vol. 32, pp. 6269, Jan. 1997.
, High-speed CMOS circuit technique, IEEE J. Solid-State Circuits, vol. 24, pp. 6270, Feb. 1989.
[44] B. E. Boser and B. A. Wooley, The design of sigmadelta modulation
analog-to-digital converters, IEEE J. Solid-State Circuits, vol. SC-23,
pp. 12981308, Dec. 1988.

Michael Choi was born in Seoul, Korea, in 1971.

He received the B.S.E.E. degree from the University
of California, Irvine, in 1994, and the M.S.E.E. degree from the University of California, Los Angeles
(UCLA), in 1998, where he is working toward the
Ph.D. degree in electrical engineering.
He is currently with Connectcom Microsystems,
Inc., Irvine, CA, developing transceivers for SONET/
SDH transmission systems at OC-192 data rate. His
research interests include analog, mixed-signal, and
data-conversion integrated circuits.

Asad A. Abidi (S75M80SM95F96) received

the B.Sc. (Hons.) degree from Imperial College,
London, U.K., in 1976 and the M.S. and Ph.D.
degrees in electrical engineering from the University
of California, Berkeley, in 1978 and 1981.
He was with Bell Laboratories, Murray Hill, NJ,
from 1981 to 1984, as a Member of Technical Staff
in the Advanced LSI Development Laboratory. Since
1985, he has been with the Electrical Engineering Department of the University of California, Los Angeles
(UCLA), where he is a Professor. He was a Visiting
Faculty Researcher at Hewlett Packard Laboratories during 1989. His research
interests are in CMOS RF design, high-speed analog integrated circuit design,
data conversion, and other techniques of analog signal processing. He served
as the Program Secretary for the International Solid-State Circuits Conference
from 1984 to 1990, and as General Chairman of the Symposium on VLSI Circuits in 1992. He was Secretary of the IEEE Solid-State Circuits Council from
1990 to 1991. From 1992 to 1995, he was Editor of the IEEE JOURNAL OF
Dr. Abidi received an IEEE Millennium Medal. He also received the 1988
TRW Award for Innovative Teaching and the 1997 IEEE Donald G. Fink Award.
He was co-recipient of the Best Paper Award at the 1995 European Solid-State
Circuits Conference, the Jack Kilby Best Student Paper Award at the 1996 International Solid-State Circuits Conference (ISSCC), the Jack Raper Award for
Outstanding Technology Directions Paper at the 1997 ISSCC, the Design Contest Award at the 1998 Design Automation Conference, and received an Honorable Mention at the 2000 Design Automation Conference.