01f53fdad01
TRANSCRIPT
-
8/8/2019 01F53FDAd01
1/4
Accuracy Limitations of Pipelined ADCs
Patrick J. Quinn
XilinxCitywest Business Campus
Logic Drive, Dublin, Ireland
Arthur H.M. van Roermund
Eindhoven University of TechnologyDepartment of Electrical Engineering
5600 MB Eindhoven, The Netherlands
Abstract In this paper, the key characteristics of the main
errors which affect the performance of a switched capacitor
pipelined ADC are presented and their effects on the ADC
transfer characteristics demonstrated. Clear and concise
relationships are developed to aid optimized design of the
pipeline ADC and error bounds are derived.
I. INTRODUCTION
Previous authors [1],[2],[3] have presented generalized anal-yses of the switched capacitor (SC) pipeline but a more compre-
hensive analysis is needed to aid its exact design. The analysis in
this paper includes, for instance, a model for capacitor mismatch,
slewing behaviour and an exact choice of capacitor sizing based
on the ADC resolution and hence noise specification.
The lowest resolution stage, using redundant-signed-digit
decoding and delivering one effective output bit, is the most
effective for achieving high throughput with high accuracy and
low power [1],[2]. The ideal transfer gain of the converter stage
is 2, while the residue equation is
, (1)
where data bits Di have values of 0 (for falling within), or -1 (for falling below ) or +1 (for
going above ). In reality, the residue voltage is affected
by a number of non-idealities associated with practical imple-
mentation which cause deviations from the ideal linear ADC
transfer characteristic. Familiarity with the effect a particular
error can have on the ADC transfer will help identify the source
of any serious errors should they arise in fabrication. The most
important errors are explored and quantified in the following sec-
tions, while the effects the errors have on the ADC transfer char-
acteristics are illustrated. Error bounds are derived to ensure the
ADC specification is met.
II. CONTEMPORARY PIPELINE ADC STAGE
The typical implementation of a SC pipeline ADC stage (1-bit
effective output) is depicted in Fig. 1 and is based on charge transfer
techniques [2],[3]. In the bottom path, the sub-ADC (just 2
comparators with thresholds set at ) reacts instanta-
neously to the analog input signal to produce a coarse digital
representation. The resultant digital code is output to a digital
decoder, while at the same time it is converted back to its
gained-up analog equivalent by the equally coarse multiplying
DAC (MDAC). The MDAC performs the level shifting, analog
multiplication by 2 and sample-and-hold buffering needed to pro-
duce the stage analog residue. On clk1, theVin is sampled on toC1
andC2 in parallel, while at the same time the sub-ADC determines
whether it falls into the top, middle or bottom of the ADC reference
range. At the end of the clk1, Vin is completely sampled on to C1
and C2, while the output of the sub-ADC is latched and held. On
clk2,C1 is switched and placed across the amplifier, closing its neg-
ative feedback loop, while at the same time only one of the input
switches ofC2 is closed by the sub-DAC using only one of clock
signals top, mid, bot, connecting C2 to either +Vref, 0, or -Vref. By
choosing C1 = C2, a nominal transfer gain of is
achieved. Note that the analysis which follows can be applied to
other implementations of MDAC, for instance, reference [2].
III. LUMPED ERROR MODEL
Each ADC pipeline stage is specified such that all accumulated
errors at the end of the conversion process remain inside 1/2LSB.
Although 1LSB is enough for monotonicity, 1/2LSB is specified
to be safe. The total fractional error is the accumulation of all
errors from theNS stages referred back to the ADC input, i.e.
, (2)
where is the net fractional stage error. Usually 1-bit redun-
dancy is employed per stage for redundant-signed digit
decoding [1] so that for a K-bit stage, K-1 bits are effectively
resolved and the stage gain factor becomes . Hence
. (3)
The magnitude of the largest allowed total input referred voltage
error is
2i iout in i ref
V V D V = -
iinV4refV iinV 4refV- iinV
4refV+
4refV
2 11 2C C+ =
Fig. 1 Contemporary single-ended SC charge transferimplementation of pipeline ADC stage (K=2)
2
2
(AnalogResidue)
C1
C2clk1
clk2
clk1
clk1
Vref
4
Vref4
Latches
and
Clkgen
top
mid
bot
Vref
Vref
top
mid
bot
bMSB bLSB
MDAC
Sub-ADC
Vini
Vout i
tot
e
1
1
i
i
j
j
NS
totGi
ee
==
=
ie
12KiG
-=
( )12
1
i
K i
NS
tot
i
ee -
==
19560-7803-8834-8/05/$20.00 2005 IEEE.
Authorized licensed use limited to: TEXAS INSTRUMENTS VIRTUAL LIBRARY. Downloaded on August 05,2010 at 05:46:57 UTC from IEEE Xplore. Restrictions apply.
-
8/8/2019 01F53FDAd01
2/4
. (4)The magnitude of the total voltage error should be less than 1/2
LSB to reliably guarantee monotonicity, i.e.
. (5)
Alternatively, each individual stage should be designed to have
a total voltage error less than 1/4 LSB of the total effective reso-
lution of the remaining stages. Note that this is approximately at
the level of the quantization noise error of the remaining stages,
which is . Assuming each stage i isK-bits, the
total fractional output error is specified as
. (6)
The stage error arises from the addition of all the errors from var-
ious static and dynamic sources. Usually, the stage is
designed such that the dynamic and static errors add up in roughly
equal measure. Hence, for the usualK=2-bit stage, the following
conditions are set for of each stage in order to ensure a
robust design:
. (7)
The various sources of static and dynamic errors are described in
the following sections.
The worst-caseINL can be evaluated from the standard devi-ation (stot) of the total equivalent input error. The error should be
obtained for the expected worst case code transition, for instance
when crosses . Assuming a Gaussian distribution for
the voltage difference between the actual and ideal transfers,
theINL can be obtained as the expected value of :
(8)
IV. LIMITATIONSON STATIC ACCURACY
A. Offset Errors
There are two basic forms of offset which have different
effects on the ADC transfer. Firstly, there is input offset which
adds up with the input signal to the stage. This offset is due mainly
to the amplifier and to a lesser extent to the switches. The transfer
function in this case is of the form:
, (9)
where the offset gets multiplied up by the stage gain. The second
form of offset is that due to the comparators which has the effect
of shifting either one or both decision levels of the sub-ADC.
The total offset from all sources must remain within the
bounds of to avoid saturating the following stage and
causing missing codes. This requirement is not dependent on the
required accuracy of the whole ADC, nor indeed accuracy of each
individual stage. The offset accumulates according to equation
(2) though, so that some low offset ADC applications may need
some kind of active offset cancellation mechanism.
In Fig. 2, the effects of input offset going beyond are
shown. The grey curves show the ideal and the black the non-ideal
transfers. Consider Fig. 2(a),(b), where the input offset goes beyond
: the ADC transfer characteristic shows a single wide code
for when the stage output voltage exceeds +Vref, followed by miss-
ing codes, the number of which depends on how much the stage out-put voltage exceeds +Vref. In Fig. 2(c),(d), the input offset goes
under , in which case the ADC characteristic first has
missing codes followed by a single wide code. In Fig. 3, the separate
effect of sub-adc comparator offsets exceeding in stage
2 are demonstrated for both the top and bottom comparators.
B. Capacitor mismatch gain errors
The generalized transfer function of the ADC stage including
capacitor mismatch errors is
, (10)
tot tot ref V Ve e
( )212 2 21
2
ref
N
N
VLSBtot ref
tot
Ve
e
=
fi
1 12 1/4LSB
( )
( )
1
1 1
214 2
1
2
ref
N K i
N K i
V
i ref
i
Ve
e
- -
- - +
fi
Fig. 2 Effects of excessive input offset on ADC transfers
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(d) ADC transfer with stage 1 offset=-5Vref/16
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(c) Stage 1 transfer with offset=-5Vref/16
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(a) Stage 1 transfer with offset=+5Vref/16
ideal
non-ideal
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(b) ADC transfer with stage 1 offset=+5Vref/16
widecode
missingcodes
Fig. 3 Effects of excessive stage 2 compr offsets on ADC transfers
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vo
ut
Vref
(c) After stage 2 with top comp offset = -3Vref/8
(d) ADC with stage 2 bott comp offset = -3Vref/8
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
out
ref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vo
ut
Vref
(b) ADC with stage 2 bott comp offset = +3Vref/8
(a) After stage 2 with bott comp offset = +3Vref/8
( )se ( )de
s d,e e
( )22
N i
s,de- - +
inV 4refV
e
e2
22
2
1
2
2
tot
tot
tot
INL E e d
e
s
p s
p
e e e
s
-
-
= =
=
( )2i i inout in off i ref
V V V D V = + -
4refV
4refV
4refV+
4refV-
4refV
( ) ( )22 1 1ci iout in i ref cV V D V D= + - + D
1957
Authorized licensed use limited to: TEXAS INSTRUMENTS VIRTUAL LIBRARY. Downloaded on August 05,2010 at 05:46:57 UTC from IEEE Xplore. Restrictions apply.
-
8/8/2019 01F53FDAd01
3/4
where represents the mismatch of nominally equal capacitors
C1 andC2. The capacitor matching requirement for the i-th 2-bit-
stage is
, (11)
An important observation with this kind of gain error is that there
is always an exact mapping of input values to output values, irre-
spective of :
. (12)
This suggests a very effective way of calibrating out capacitor
mismatch errors by extrapolating between these correctly
mapped points.
The effects of capacitor mismatch errors in stage 1 are illus-
trated in Fig. 4 Positive capacitor mismatch errors (Fig. 4(a),(b))
cause a series of wide codes and negative code jumps resulting
in non-monotonicity. Negative capacitor mismatch errors (Fig.
4(c),(d)), on the other hand, cause a series of narrow codes and
positive code jumps, resulting in missing codes. Fig. 5(a),(b)
show the result of positive capacitor mismatch in stage 2. Here
again there are wide codes corresponding to the six transition
points from the ADC input to the second stage output. Further-
more, there are also two sets of missing codes at which
are the transition points of the previous stage. A similar scenario
exists for negative capacitor ratio errors (Fig. 5(c),(d)), except
that wide codes are replaced by narrow codes and negative and
positive code jumps are interchanged.
C. Amplifier Gain Errors
The OTA, in a switched-capacitor configuration of any stage
i along the pipeline, must linearly amplify the input voltage by a
factor of 2 over the full scale range of -Vref to +Vrefto within amaximum error of . As a result of the sampled-data oper-
ation of the stage, only end values count at the end of sample clock
periods. Hence, only DC gain variations around count,
since the differential OTA input always settles back to around 0V
as the output end value is reached. The OTA DC gain has an effec-
tively even characteristic and the DC gain doesnt vary much
around whether the outputs end up at 0V or at .
The stage transfer, including OTA gain error , is
, (13)
. (14)
In general, the DC gain requirement for stage i is given by
. (15)
To take an example, say the SC ADC stage of Fig. 1 is
designed as the first stage in a 12-bit pipeline ADC. The feedback
factor is . With nominally, , so
that the amplifier DC gain is required to be !
An important observation with this kind of gain error is that
there is always an exact mapping of the zero crossings irrespec-tive of DC gain error, be it linear or non-linear, i.e.
.
IV. LIMITATIONSON DYNAMIC ACCURACY
A. Linear and Non-linear Settling Constraints
The stage amplifier slews when the step voltage at the ampli-
fier differential input goes beyond the maximum linear input
range of which corresponds to it delivering its maximum
currentIto the load. The dynamic settling error caused by
the amplifier not settling out in sample period Tis
(16)
. (17)
Note is that proportion ofVstep which the amplifier
input sees through capacitor division before it starts to react. This
error (16) must be no larger than the required stage dynamic set-
tling error, . The effect of combined finite set-
tling time and finite slew rate is shown in Fig. 6. The limited
settling time produces a similar effect to limited DC gain in the
sense that end values are not achieved within a sampling clock
cD
1
1
3 2N iC
s - -D
Fi . 4 Effects of sta e 1 ca mismatch errors on ADC transfers
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(a) Stage 1 transfer with positive c (c) Stage 1 transfer with negative c
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(b) ADC transfer with stage 1 positive c (d) ADC transfer with stage 1 negative c
allwide
code
s
missingcodes
cD
{ } { }0 0in ref ref out ref ref V V , , V V V , , V - + - +
4refV
( )22 N i- - +
0inVD =
0inVD = refV
Fig. 5 Effects of stage 2 cap mismatch errors on ADC transfers
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(a) After stage 2 with positive c in stage 2 (c) After stage 2 with negative c in stage 2
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(b) ADC transfer with stage 2 positivec (d) ADC transfer with stage 2 negative c
missing codes atstage 1 transitions
0Ae
( ) ( )0
2 1i iout in i ref A
V V D V e= - +
0 0
1wherefb
A A be -
210 2
fb
N iAb
- +>
1
1 2
C fb C C b += 1 2C C= 12fbb =
( )0 dB 84dBA >
{ }4 40 0ref ref V VinV , , - +
2 onVsettlee
2
2 2
,
,
on
on on
step
Vnstep
settleV Vn
stepV
e V
e V
g
g g
e-
-
= >
( )2here and 1stepslew onVT t
slew Vn t
g
tt
-
= = -
Vstep
( )22
N i
settlee- - +
1958
Authorized licensed use limited to: TEXAS INSTRUMENTS VIRTUAL LIBRARY. Downloaded on August 05,2010 at 05:46:57 UTC from IEEE Xplore. Restrictions apply.
-
8/8/2019 01F53FDAd01
4/4
interval. This is a linear effect though, unlike limited slew rate
which is input level dependent. This is clear from Fig. 6, where
large input levels up to cause a tapering off effect leading
to poor INL and harmonic distortion. Note again that the zero
crossings are correctly mapped despite the non-linear distortion.
B. Thermal Noise
The output noise power produced by a single ADC stage,
such as Fig. 1, has 3 constituents:
, (18)
with the amplifier (OTA) noise sampled only in the write
phase and the noise produced by the switches in
both the read and write phases. The thermal noise power of the
OTA, referred to the output, is given by
, (19)
where is the noise power spectral density of a differ-
ential pair and is the noise excess factor depending on OTA
architecture used. The amplifier input noise power is multiplied
by , which is the noise power gain from stage input
to output. For an essentially first order amplifier settling
response, the amplifier noise bandwidth is
, (20)
with the linear settling time constant and the total
effective load capacitance. Hence, the amplifier noise power for-
mula can be reduced to
. (21)
For the SC ADC stage, the switch noise contributions for the read
and write phases can be calculated as:
. (22)
The factor 2 in (22) is for the differential case and
is the bandwidth reduction factor in the write phase Since noise
in the read phase is pre-sampled, it is not affected by the OTA band-
width. Generally, and withC1=C2, the ADC stage
output noise power for full Nyquist operation becomes:
. (23)
The equivalent input noise power for each stage is obtained by
dividing the stage output noise power by the signal gain up to that
stage. For similar stages, the power gain factor per stage is just
, and the total equivalent noise calculated back to the input is
.(24)
Substituting (23) into (24), the total input-referred pipeline ADC
noise power ends up as
. (25)
To ensure that the totalSNR is degraded by just 1.76 dB, the ther-
mal noise power is specified with respect to the quantization
noise power by . Hence,
. (26)
The minimum value ofC, based on noise considerations, can
be found for the SC MDAC pipeline stage of Fig. 1. Assuming
parasitic capacitance at the amplifier input is small compared
toC, then , which with ,
gives . is typically in the range of 1 (single-
stage OTA) to 2 (dual-stage OTA) - is chosen here.
The minimum value ofCis now defined by:
. (27)
For a 12-bit differential application with 1V signal range and full
Nyquist operation, . Note the importance of maximiz-
ing signal range for the sake of minimizing signal capacitance, sothat if the signal range is doubled to 2V, the capacitance is reduced
by a factor of 4 (C=3/4pF) for the same 12-bit performance.
III. CONCLUSIONS
A lumped error model was developed in this paper to account
for static and dynamic errors due to hardware imperfections in
fabricated pipelined ADCs. Errors arising, for instance, due to
imperfect capacitor matching, non-linear settling behavior and
noise were analyzed. The effects of various expected error
sources on the ADC transfer characteristics were described. It
was emphasized that familiarity with the effects specific errors
have on the ADC charactersictics can be a useful aid for debug-
ging hardware errors after ADC fabrication.
REFERENCES
[1] S. Lewis, Optimizing the stage resolution in pipelined, multi-stage,
analog-to-digital converters for video-rate applications, IEEE
Trans. Circuits Syst. II, vol. 39, Aug. 1992.
[2] P. Quinn, M. Pribytko, A. van Roermund, Calibration-Free High-
Resolution Low-Power Algorithmic and Pipelined AD Conversion,
in Analog Circuit Design, Kluwer Academic Publishers, ISBN 1-
4020-2786-9 (HB), 2004, pp.327-349.
[3] Thomas B. Cho and Paul R. Gray, A 10 b, 20 Msample/s, 35mW
Pipeline A/D Converter, IEEE J. Solid-State Circuits, vol. 30, no.
3, pp. 166-172, Mar, 1995.
[4] J. Goes, J. Vital, and J. Franca, Systematic design for optimization
of high-speed self-calibrated pipelined A/D converters, IEEE
Trans. Circuits Syst. II, vol. 45, Dec. 1998.
refV
-0.5-0.75-1 -0.25 0 0 .25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(a) Stage 1 transfer with slewing in 1st stage (c) After stage 2 with slewing in 2nd stage
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
-0.5-0.75-1 -0.25 0 0.25 0.5 0.75 1
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
Vin Vref
Vout
Vref
(b) ADC transfer with slewing error in 1st stage (d) ADC transfer with slewing error in 2nd stage
Fig. 6 Effects of non-linear settling errors on ADC transfers
2 2 2 2
stage OTA C C read write N N N N s s s s = + +
2
ampNs
2 2,
C Cread writeN Ns s
( ) 22 16 1 13 1OTA OTA OTAm fb N e N gkT N B bs = + 16 13 mg
kT
OTAeN
2 21 fbG b=
14 4
fb m
OTA OTA Leff
g
N CB
b
t
=
OTA
tLeff
C
( )2 43 1OTA OTA fb LeffkT
N eCN
bs = +
( )21 2 1 2
22 22 andNOTA
C C Nread writefb swit
BkT kT kT N NC C B C C b
s s
+= = +
OTA switN NB B
OTA swit
N NB B
( )22 1 43 1stage OTALefffbkT kT
N fb eC CN
bs b + +
21 fbb
( )2
2
2 2 2 2 4 2
11
fb
tot stage stagefb
N N fb fb fb N b
bs s b b b s
-= + + +
( )22 1 431 1tot OTALefffbkT kT
N fb eC CN
bs b
- + +
2sD2 21
2totNs sD<
( )2
2 2
1 4 13 241 2
1NOTALefffb
kT kT FS fb eC C
Nb
b-
+ + <
1 2 2 L fbeff
C C C C b= + + 1 2C C C= =52Leff
C C= OTAe
N
2OTAe
N =
2
2
245N
FSC kT>
3 1pFC .>
1959
Authorized licensed use limited to: TEXAS INSTRUMENTS VIRTUAL LIBRARY. Downloaded on August 05,2010 at 05:46:57 UTC from IEEE Xplore. Restrictions apply.