Documente Academic
Documente Profesional
Documente Cultură
I. INTRODUCTION V in
+
2
M
Vout
_
The design of high-performance analog-to-digital convert-
ers (ADCs) presents difficult challenges as applications call
for higher speeds and resolutions and as device dimensions Sub Sub
ADC DAC
and supply voltages are scaled down. To meet these chal-
lenges, increasingly sophisticated ADC architectures, circuit M Bits
MDAC
Equations (2)-(5) do not lend themselves to a closed-form
2
solution but are readily solved in MATLAB using the func-
1 C2 1 tion roots. That is, the output of each stage in the pipeline
is calculated based on the user-specified nonidealities of the
1 C1
VX
stage and passed on to the next stage.
V in
A1 Vout The coefficients j in Eq. (1) are obtained by transistor-
CX
CL level simulation of the op amp and fitting a third-order
2
polynomial to the resulting characteristic. For differential
D1 implementations with small mismatches, 2 0, thus
KVREF simplifying the fitting procedure.
The offset of the comparators is incorporated by adjusting
V REF D1 K their decision thresholds and hence producing a correspond-
4 00 1 ing digital output for the sub-ADC. The offsets can be
01 0 specified by the user for all stages.
V REF
+ 11 +1
4 While the above modeling procedure has been derived for
SubADC a 1.5-bit stage, our simulation tool in fact handles arbitrary
resolutions in each stage of the pipeline. Equations (2)-(5)
Fig. 2. Pipelined stage with 1.5-bit resolution. can be modified to handle N -bit MDACs with or without
redundancy.
role.
B. Modeling of Dynamic Effects
In the analysis and design of pipelined ADCs, it is desir-
A. Modeling of Static Effects able to quantify how the precision degrades as the clock rate
We now formulate the behavior of the stage so as to capture increases. The simulation tool developed here approximates
the above nonidealities. The input-output static characteristic the MDAC dynamic behavior by a slewing regime followed
of the op amp in Fig. 2 can be approximated by a third-order by a linear settling regime . The time at which the settling
polynomial across the voltage range of interest: changes from nonlinear to linear (t1 ) is a user-defined fraction
of the clock period. The linear settling behavior is based on
Vout 1 Vx + 2 Vx2 + 3 Vx3 . (1)
the small-signal model shown in Fig. 3, whose time constant
In the sampling mode, the input and output voltages of the is given by [3]
op amp are equal and expressed by CL (CS + Cx ) + CL CF + (CS + Cx )CF
amp = . (6)
2
Vx1 = 1 Vx1 + 2 Vx1 3
+ 3 Vx1 , (2) Gm CF
In other words, an exponential with this time constant and
where Vx1 denotes Vx in the sampling mode. (The offset of
an initial value equal to Vout (t1 ) is solved to obtain Vout at
the op amp can also be included by subtracting it from Vx1
t = Tck /2.
on the right-hand side.) The total charge stored on C1 and
C2 is thus equal to CF
2 3
Vout = 1 Vx2 + 2 Vx2 + 3 Vx2 , (4)
where Vx2 is the value of Vx in the amplification mode. Also C. Modeling of Noise Transfer Functions
the total charge residing on C1 , C2 and Cx in this mode is Perhaps the most difficult task in the analysis of pipelined
given by ADCs, computation of the overall input-referred noise plays
a critical role in the power optimization of the design. Since
Qamp = C1 KVref + C2 Vout (C1 + C2 + Cx )Vx2 . (5)
such ADCs typically scale down the devices and bias currents
Conservation of charge requires that Qamp = Qsamp . Equa- along the pipeline, it is essential to accurately calculate the
tions (2)-(5) provide a complete description of the static noise contribution of each stage.
behavior of the pipelined stage, modeling the following The difficulty in noise computation arises from two fac-
nonidealities: finite op amp gain, op amp nonlinearity, op tors: (1) the noise transfer functions become very complex,
amp offset, op amp input capacitance, and mismatch between assuming orders as high as 6 or 7 (explained below), and
C1 and C2 . (2) the noise sampled by the input of one stage depends on
403
3
bandwidth limitations imposed by the MDAC in the previous sources included in the previous calculations. The state space
stage. Cadences Spectre can calculate noise in transistor- approach can be extended to this case as well.
level transient simulations, but the simulation becomes pro- The foregoing noise calculation method continues to the
hibitively long for a cascade of pipelined stages and, owing end of the pipeline, thereby yielding the total integrated noise
to difficulty in finding periodicity, fails in some cases. In this at the output of the last MDAC stage. The result is then
work, we apply the state space method of analyzing linear divided by the total gain provided by the pipeline so as to
systems [4] to the calculation of the noise. arrive at the input-referred noise voltage of the ADC.
Figure 4 shows the small-signal model of an MDAC The circuits of Figs. 4 and 5 may imply that the value of
employing a two-stage op ampa common choice in todays the switch resistances must be provided by the user. Since
low-voltage ADCs. Here, Rsj and Vsj respectively denote the these resistances should not limit the settling behavior of the
on-resistance and thermal noise of the switches that appear MDAC, the tool assumes that the resistance in series with
in series with the sampling capacitors, and In1,op and In2,op each capacitor leads to a time constant of 1/10 of Tck /2.
the noise current at the output of each stage of the op amp Also, the op amp noise spectral density is approximated
(per unit bandwidth). The op amp is modeled by two gain by 4kT gm where denotes the excess noise factor of
stages and a compensation capacitor Cc . Note that the input MOSFETs ( 1) and gm denotes the transconductance of
sampling network of the next stage is also included. The each stage of the op amp. As explained in Section V,
objective is to determine the total noise sampled on CL at these approximations lead to good agreement between our
the end of this mode. simulators estimate and that obtained from Spectre.
Rather than directly compute the noise transfer functions,
we utilize the state space method: we first assign independent III. S IMULATOR E NGINE
states to the storage elements in the circuit (voltages Xj in
The simulator engine is shown in Fig. 6. The arguments
Fig. 4) and then write the following matrix equations:
to the engine are the MDAC architecture, number of stages
X = AX + BU (7) in the pipeline, clock frequency, input signal swing, input
Y = CX + DU , (8) frequency and the variance of the clock jitter. The INL
and DNL profiles are obtained by code density tests using
where X denotes the states, X = dX/dt, U is the vector either a ramp input or a sinusoidal input. The ramp input is
of inputs to the system (the noise sources), and Y is the chosen such that it produces 10 codes per bin. For sinusoidal
vector of the outputs (in this case, the voltage across CL ). inputs, the computation becomes difficult as the resolution
Also, A, B, C, and D are the state matrices. exceeds 8 bits because hundreds of thousands of samples
The circuit of Fig. 4 contains six states, five inputs and one are necessary to produce the profiles. Thus, the simulation
output. Since the output is equal to state X6 , we have D = 0 is performed in a number of passes, each generating 25,000
and C = [0 0 0 0 0 1]. Matrices A and B are defined by codes. The results are then combined properly, yielding the
the circuit parameters and can be readily found by writing DNL and INL profiles.
the various nodal equations. For example, the first row of A, In order to determine the output spectrum in response to
i.e., a11 , ..., a16 , is obtained by noting that an input sinusoid whose amplitude and frequency are defined
by the user, the simulator first generates the digital bits
X 1 = a11 X1 + + a16 X6 + b11 Vn1 + from each stage, appropriately combines them to generate the
+b15 In1,op + b16 In2,op , (9) digital representation of the analog value, and subsequently
takes the FFT in MATLAB.
where b1j are the first row of B, Vnj are the four noise
sources, and Inj,op the op amp noise currents. To compute
a11 , we apply a KVL in the loop consisting of C1 , CX , Rs1 , IV. S IMULATOR I NTERFACE
and Vns1 , thus arriving at a11 = 1/(Rs1 C1 ). With the four The simulator has been implemented in MATLAB and
state matrices known, Eqs. (7) and (8) are numerically solved subsequently compiled to an executable that can be run on
to find the contribution of each noise source to the output. various platforms. On the GUI, the user selects the resolution
These contributions are then integrated for a frequency range of each stage (1 bit, 1.5 bits, or higher integer values) and the
of 0 to and summed to determine the total noise. The number of stages in the pipeline. The user can also enter the
numerical integration is simplified through the use of initial following parameters: (1) for each stage, the op amp gain and
value theorem in Laplace transforms. nonlinearity coefficients, the capacitor values and mismatch,
The circuit of Fig. 4 still does not completely represent the the op amp and comparator offsets, and the transconductance
noise mechanisms in a cascade of pipelined stages. Consider of the op amp; (2) the analog signal swing and frequency,
the cascade depicted in Fig. 5, where the first stage is in the and the clock frequency.
amplification mode and the second in the sampling mode. Upon simulation, the user can view the residue plot of
Here, the second op amp is configured as a unity-gain buffer each stage, the overall input-output characteristic, the DNL
by a feedback switch with noise Vnf 1,2 . Thus, the noise of and INL profiles, and the output spectrum. The total input-
the op amp and the feedback switch is sampled on C1,2 referred noise along with contributions from each stage are
and C2,2 in the second stage along with the other noise also computed.
404
4
2 Next Stage
V nS2
C2 R S2 2
V nS3
X3 R3
CC
C1 X6 CL
R4
X1
R S1 X2 CX g m1V 1 I n1,op
2 V1 R1 X4 C O1 V 2 g m2 V 2 I n2,op R2 X 5 C O2 2
V nS1 V nS4
Op Amp
MDAC 1 MDAC 2
Fig. 5. Noise sources: stage 1 in amplification mode and stage 2 in sampling mode.
Solve
Residue
Eqs. (2) (5)
Input for Residue Plot Residue Plot
for Stage j
I/O Plot and DNL/INL Plot
Ramp
Stage j
Generator Combine
Digital Output Bits
I/O Plot
The simulator also provides two calibration algorithms, V. S IMULATION R ESULTS AND
either one of which can be selected and run on the ADC.
Alternatively, the user can use his/her own calibration al-
C OMPARISON WITH S PECTRE
gorithm and compare the performance of the ADC before
In order to demonstrate the speed advantages of the
and after calibration. In the latter case, the user can write
proposed simulator, pipelined ADCs having resolutions of
the digital outputs of the stages to a file, run the calibration
8 to 12 bits have been simulated by the tool, by Spectre, and
algorithm, and return the results to the simulator so as to
by Ultrasim on a 64-bit AMD, 2.2-GHz machine with Redhat
study the performance after calibration.
Linux. Spectre and Ultrasim simulations were performed at
transistor level (in 90-nm CMOS technology).
Table I compares the linearity simulation times using our
405
5
DNL (LSB)
Resolution Ultrasim Proposed Simulator 0.05
8-bit 5h:33m:40s 3.3s 0.1
10-bit 12h:45m:33s 45.97s
0.15
12-bit 63h:11m:24s 16m:37s
0.2
0.25
simulation time to compute the 1024-point FFT of the output 200 400 600 800 1000
Code
for 8-bit, 10-bit and 12-bit ADCs. Table III compares the
TABLE II (a)
1024 FFT COMPUTATION TIME
Resolution Spectre Ultrasim Proposed
Simulator 1
INL Plot
INL (LSB)
Accuracy tightened for 12-bit resolution.
0
0.05
Stage 2 197 208.99 6.1% 4m:36s 0.192s 0.1
Cascade 275 299 8.7% 26m:18s 0.3s 0.15
0.2
0.25
0.3
R EFERENCES 0
406