742

IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. SC-17, NO. 4, AUGUST 1982 F. F. Tsui, JSP-A research signal processor in Josephson technology, IBM J Res. Develop., vol. 24, pp. 243-2523 Mar. 1980. T. R. Gheewata, Design of 2.5 micrometer Josephson current injection logic (CIL), IBM J. Res. Develop., vol. 24, pp. 130142, Mar. 1980. P. C. Amett and D. J. Herretl, Regulated AC power for Josephson interferometer latching logic circuits, IEEE Trans. Magn., pp. 554-557, Jan. 1979. M. B. Ketchen, Power supplY regulators for Josephson latching logic, in Proc. IEEE Int. Confi Circuits and Comput., ICCC80, Oct. 1980, pp. 874-877. D. J. Herrell and N. Raver, The electrical characteristics of a high performance package for Josephson technology, presented at the High Speed Digital Technol. Conf., San Diego, CA, Jan. 1980. H. C. Jones and D. J. Herrelt, The characteristics of chip-to-chip signat propagation in a package suitable for superconducting circuits, IBMJ. Res. Develop., vol. 24, pp. 172-177, Mar. 1980. S. K. Lahiri, P. Geldermans, G. Kolb, J. Sokoloski, and M. J. Patmer, Pluggable contacts for Josephson packaging, J. Eletrothem. Sot. Extend. Abstr., vol. 80, no. 1, p. 216, 1980. J. H. Greiner et al., Fabrication process for Josephson integrated circuits, IBh!l J. Res. Develop., vol. 24, pp. 195-205, Mar. 1980. [11 ] D. E. McCumber, Effect of ac impedance on dc voitagecurrent characteristics of superconducting weal-link junctions, J. APP1. Rhys., vol. 39, PP. 3113 -3118, June 1968.

[3] [4]

[5]

[6]

[7]

[8]

[9]

[10]

Melvin Klein (A42-M53) was born in New York, NY, in 1920, He received the M.E. and M.S. degrees from Stevens Institute of Technology, Hoboken, NJ, in 1940 and 1942, respectively, and the M.A. degree from Columbia University, New York, NY, in 1948. He worked at IT&T on electronic navigation equipment from 1942 to 1946, and at the Radio Receptor Company from 1950 to 1958 on radar equipment and semiconductor devices. In 1958 he joined IBM and worked in the Com~onents Division on semiconductor devices for logic and memory appli~ations. He was Manager of Semiconductor Device Development and Packaging for Magnetic Memories from 1966 to 1969. In 1973 he became a Research Staff Member at the IBM Research Center, Yorktown Heights, NY, where he is engaged in the development of Josephson devices and circuits for high-performance computer applications.

Noise Sources and Calculation Techniques Switched Capacitor Filters

JONATHAN H. FISCHER

for

Absmact-The noise response of switched capacitor networks (SCNS) is reviewed with emphasis on simplifying approximations suitable for SPICE noise simulation. The techniques developed cover all op-arnp noise sources, as wetl as capacitor switching noise, The close agreement between predicted and measured noise responses for several monolithic SCNS bears out the vstklity of these simulation techniques.

()+- ()s()
S(t)= ~8(t-kT)

Fig, 1. Ideal sampler.

I. INTRODUCTION

HE noise response derivations for switched capacitor networks (SCNS) developed in this paper differ from earlier work [1] - [3] by emphasizing approxitrrations that facilitate the use of general-purpose programs, such as SPICE, for accurate SCN noise analysis. The derivations will cover ideal sampling effects, the development of a suitable SCN integrator noise model, and computer simulation techniques. Applying the noise model to practical SCNS shows that the relative contributions of op-amp 1/f, foldover flat-band, and capacitor switching noise are filter topology dependent. II. MODULATION AND FOLDOVER
EFFECTS

rate (f~). The baseband (~n) and sideband (~~b) spectral densities of Fig. 2 are equal as a result of ideal impulse sampling [4], [5]. A convention that will prove useful in the analysis to follow is to number the sidebands so as to associate each sideband with the sampling frequency harmonic it is centered about, as in Fig. 2. For the more general case where the signal bandwidth is greater than f~/2 (such as the output noise of an op-amp in an SCN), aliasing will occur. As an example, assume that
the sigmd to be sampled has a bandwidth the sample rate. (Z3Wn) of three times

To start the analysis of noise in sampled data systems, we begin with the effects of the idealized sampling operation of Fig. 1 on a signal band limited to less than half the sample
Manuscript received May 6, 1981; revised January 7, 1982. The author is with Belt Laboratories, Hohndel, NJ 07733.

The first five sidebands resulting from the sampling operation have been depicted in Fig. 3(a). Referring to the figure, the following are contributors to the frequency band from dc to f.: the fundamental, *1, +2, and the +3 sidebands. The addition of these sidebands is depicted in the stacked structure of Fig. 3(b). This spectral stacking of an undersampled signal results in an output of larger spectral density than the input for the frequency band from dc to f8/2. The noise gain of a track-and-hold circuit is shown in Fig. 4 (and is discussed in Section III). 01982 IEEE

0018 -9200/82/0800

-0742 $00.75

FISCHER: NOISE SOURCES AND SWITCHED CAPACITOR

FILTERS

743

F(f) I FUND -fin / f fin

o (a)

-~
s a s

Fig. 5. Integrator topology under study.

(b)
-2 -2f, -1 -f* FuND o (c) +1 fs +2 f 2f,

-+mvo
1-

Fig. 2. (a) Input spectrum.

(b) Sampling function. signal spectrum.

(c) Modulated

!al-r-L,
TRACK %OLO T #T~

I
1 -i J-2 -2 -2 -4 -2 -1 -4 +3 ,+1 -2

FU,ND -1 +1 -1 -1 [ +j +4 ,+1 +2, +2

I ,+4 +2

~I +2 4~-

-2 -2-21 I +2 t2 +2 -3 I 1+3,

b=
H SA~p$EAR CT

T
t

-~
T

SH SH - ~

Fig. 6. Track-and-hold circuit.

1-3 -4-4-4 -5 -5

-3

-3 -4

-3

-3-3 :4-4 -5-51

+3, .,. I +4,+4

+3,

+3. ,+4

,,+3,
,+4

~ ,+.4 +5 ,+4: ,+56fs + ON OFF ON tT _f_ T OUT

-5

-5

+5 +,5 +5, +5 -6fs -5fs -4fs -3fs -2fs -fs O (a) f: 2fs 3;s 4;s

5;s

~E
-13 -8 -2 -2 -2

-2

-2

-2

-2

+4

+4

+4

+4

+4 &

SAMPLE OFF CLEAR o: E-

tH4

t t

-6fs -5fs -4fs -3fs -2fs

-fs

fs

2fs

3fs

4fs

5fs

6fs

(b) Fig. 3. Impulse sampling (a) sidebands and (b) resultant frequency spectrum.

Fig. 7. Track-and-hold equivalent circuit.

()
Y Y T/H

IN dB OUTPUT

INPUT NOISE

where (2BWn/f,) - 1 is the number of sidebands falling in the frequency range dc < f < f.. To simplify the calculations, BWn/f, will be treated as an integer. If a fraction of a sideband folds into band, BWn/f, is incremented to the next larger integer to give an upper bound on the in-band noise. Recalling that ~,1, = qn simplifies (1) to (2) III. NOISE MODEL OF A TRACK-AND-H
L,D

BWn = 400 KHz f, = 100 Hz TT,H= 0.5

Fig. 4. T)H noise response. For the remainder of this paper, IllVn represents the equivalent noise bandwidth of a noise source, and will be taken to be that bandwidth required to contain the same noise power as the source, but with a uniform spectral density rIn. If the input is a white noise source, the frequency-shifted sidebands are now uncorrelated with the fundamental or each other; hence, power rather than voltages are added to compute the total output density (qT). Putting these results into mathematics for .f~< .f< f~,

One of the basic building blocks commonly used in SCNS is the integrator of Fig. 5. This integrator is a form of a While the track operation is track-and-hold (T/fI) circuit. not inherent to all SCN integrator topologies [6] - [8], this property will be included for completeness. The integrator analysis will begin by describing transmission gate (input tracking) and S/Ii (hold operation) noise responses using the T/H of Fig. 6. These results will be experimentally verified in Section IV, and a complete integrator model will be developed in Section VI. The T/H can be viewed as two systems in parallel, as in Fig.
7, The transmission gate models the feedthrough operation The sample-and-hold (S/n) when the T/Ii tracks the input.

=on+w?ws

(1)

models the hold operation by sampling at the instant the transmission gate opens, while resetting to zero during the track mode.

744

IEEE JOURNAL

OF SOLID-STATE

CIRCUITS, VOL. SC-17, NO, 4, AUGUST

1982

A.

Track Model
IN+

If the input is white noise of bandwidth B Wn, the results of Appendix A can be used to obtain the results below relating
the output

to the input power density:

noutm

= ~% *
TT~n

BWn <:

f~

(3a)
Fig. 8. T/H frequency domain model in terms of S/H duty cycle (.~H)

BWn > 10f$

(3b)

for Bwn > loffi

where t-~ is the track time/sample period duty cycle. Note that the output spectral density (in the passband) is never greater than that of the input (for a white noise source).

% ~ (,) INdB
T/H OUTPUT+ INPUT NOISE -

B. Sample-and-Hold

Noise Model

and then holds for An ideal S/H samples instantaneously, the remainder of the sampling period. The hold time is the S/H duty cycle times the sample period (r~~ X T), as in Fig. 7. Using the results of Appendix B for a white noise input, Tout(f) ~ mkk7 TSHf sinc2

Hz 0 200 400 600 800 1000

()
f.

BW. = iO KHz

BW. < ; f.
(4a)

f, = IOHZ TT/H= 0.5

Fig. 9. SCN noise floor.

ou(f)<2n(?)(TH)

sinc2 (%)

Wn>+fs
(4b)

Case 2) r~~ = l. Ideal S/H results, and the track term drops, leaving the familiar sine (x) envelope.
Tout(f) < zf?n~~fi (~)sincz[~].

Thus, foldover effects of ideal sampling followed by a hold

operation (that forms the ~/~) can produce an in-band response that is significantly larger than the input spectral density.

case 3) TSH = 1/2

and BW. > 10f~:

C Frequency Domain T/H Noise Model

Since the transmission gate and the S/H operate during nonoverlapping time intervals and the noise is white in nature, their power spectra simply add [3], [4] . The track-and-hold operation occupies the entire sampling period: rT+rsH=l. Recasting the transmission gate response in terms of 7SH and combining with the S/H results, the frequency domain model of Fig. 8 can be constructed, Writing qout(f) as a function of ~n for wide-band white noise, Vout(f)

~out(f) =O.

[X-%sinc(ik)+l
(*)+~Vn.

Case 4: To investigate the out-of-band response, hold f8 and 7SH constant (let TSH = 0.5 for simplicity), and increase f:
nout(fl=nn&)sinc2

Recalling that \sinc (x)1 < I/lxl,

n(%)

(%2

= [2
~n ~ V.
Z7SH

TsH Smc

2[%1+( 1

-4
(5) As f increases, the envelope of t?o~t(~) decreases with the square of .f, sofor f >> f$ [see (5) and (6)], the transmission gate term sets the noise floor in Figs. 4 and 9 at ~n/2.
IV. EXPERIMENTAL OF FOLDOVER VERIFICATION THEORY

BWn < + f$

~out(f)

(i+inz[%l-d
(6)
more, the

BWn > lof~.

To gain physical insight into the above results, let us look at a few limiting cases. Case 1) rs~ =0: The tracking gate is always closed, hence no sampling, thus the S/H term drops and ~Out( f ) = Vn.

To aid in the isolation of various noise sources, the test circuit of Fig. 10 was used. To eliminate any effects resulting from finite bandwidth of the op-amps or switches, the noise source was selected to produce band-limited white noise with the cutoff frequency far below fT of the op-amp. Furtherinput noise level was set as far above the noise

FISCHER: NOISE SOURCES AND SWITCHED CAPACITOR

FILTERS

74s

S*
WHITE NOISE + GENERATOR S4

( ,)
~
Vo

lNdB

INPUT

OUTPUT

CLOCK GENERATOR
I 1

1
w CMOS 140666) (MG TRANSMISSION GATE

. 1
IOK 1$

I IN KHz

Fig. 10. Transmission gate and S/H

test

circuit.

TABLE I (a) TRANSMISSION GATENOISERESPONSE ERSUS V DUTYCYCLE (f<< FORBWn = 100 kHz AND ~, = 1 kHz. (b) S/H RESPONSE(f<< FOR BWn = 100 kHz AND ~, = 1 kHz.
(a)

Fig. 11. Transmission gate wide-band input regponse.

J) f,)

; ()

INdE

OUTPUT

INPUT

95
61 50 35

o
-1.8 -2.s -4.2 -7.0 -13

-0.2
-2.1 -3.0 -4.6 -7.4 -13

I
Duty Cycle

1s
5

I
Calculated

f IN KHz .
01

Z3456T

8940

BWn =IOOKH, f,= 4 KHz 7$,

(b)
Output referred to input Measured (dB) Calculated First null frequency Measured

=0.50

Fig. 12. S/H widbband input response. TABLE H 50 PERCENT) FOR (a) Ill?. = 100 kHz (b) BWn = 10 kHz
=

(dB) 22.6 21 17 14 9.4 3.8 -3

(kHz) 1.0 1.2 1.s 2.s 4.5 8.8 19

(kHz) 1.0 I .4 2.0 2.9 4.s 9,1 20

95 80 50 34 21 11 5

20 20 15 12 9 4 -3

T/H

CIRCUIT RESULTS (TSH

AND

(a)
f.
(Hz)

Predicted Measured
density density

floor as was practical without circuit limiting. The unstitched (switch latched in closed position) noise signal level serves as the reference for all the results.

M
lOOk 10k lk 100 (dB) -0.2 7.4 17 27

(dB) -2 7.4 Is 29

A. Track Model Verification

Equation independent 10f~), and calculated marized in in Fig. 11. (3) indicates that the track noise contribution is of noise bandwidth or sampIe rate (for BWn > is dependent only on the track duty cycle. The and measured track operation results are sumTable I(a) with the frequency independence shown
The

z
(Hz) 10Ok 10k Ik 100 10

(dB) -3 o 7.4 17 21

(dB) -3.8 1.5 8.5 17 26

S/H

response

extends

below

the input noise level because

the track operation, with its associated noise contribution, not present in the S/H.

is

C. T/H Noise Model Verification

To investigate the foldover term and the orthogonality of the track and S/H operations, the T/H of Fig. 6 has been utilized with a clock duty cycle of 0.5 (50 percent). Sample Noise Calculation: Calculating the dc spectral density for the reduced sample rate off. = 100 Hz (measured response
in Fig. 4),

B. S/HiVoise Model Verification Using (4b) to predict the S/H

wide-band input,

response

at dc (v~~a.)

to a

ma = 2T;H

()
f,

BWn

BWn = 100 kliz,f~ = 100 Hz.

Using (6),

For f~ = 1 kHz, BWn = 100 kHz, and rs~ = 0.5, = 50qn (17 dB gain). First null occurs at 1 kHz/O.5 = 2 kHz. The white noise response is shown in Fig. 12, with similarly calculated and measured results summarized in Table I(b).

= 500 (27 dB gain).

Similar results are to be found in Table II.

746

IEEE JOURNAL

OF SOLID-STATE

CIRCUITS, VOL. SC-17, NO. 4, AUGUST

1982

N+l--m+izl--EOuTuT
(a)

$E1-HEH-EF+zl-Ol
(b) (a)

(c)

Fig. 13, Cascaded T/Hs with in-phase (a) and staggered (b) phase clocking. The resampled wide-band noise spectrum for the staggered clock case is shown in (c).

Fig. 14. (a) Charge accumulator and (b) simplified accumulator with
V~ngrounded.

V. EXTENSION

TO CASCADED

STAGES

There are two extreme cases when cascading T/H stages. One is when all the T/Hs have the same timing (in-phase
clocks) as in Fig. 13(a). The other extreme and after 13(b). clock is when one

T/H

tracks, and the T/H before

(staggered might [6], siderations [7], clocks) both make [11]. as in Fig. clocking contain

it are in the hold state While SCN parts of an SCN realization con-

The topology of the charge accumulator (with the l/~ and flatband op-amp noise modeled by Fn) of Fig. 14 is representative of a basic structure used in SCN filters [6] - [8], so a careful analysis is in order.

schemes,

A. ChargeAccumulator Model
To aid in visualization, the accumulator is redrawn in Fig. 14(b) with Vk grounded. Assume as initial conditions that qcr = qc~ = O and that the op-amp voltage gain is infinite. After switch SI has closed, the feedback action of CI will force the output to Vo=
(while

the staggered

case the most common

In one extreme, all the T/H stages are driven by the same Because the significant foldover effects are clock phases. present only in the hold phase (assuming a memoryless track phase), the noise at the T/H input is simply the sum of the
unsampled noise feeding through

to

the stage

of

interest.

When

the

T/H changes to the hold phase, the input noise is

()
1+%
not from previous

Vn
input to equal Vn) and to track Vn

sampled and folded down to baseband. The resulting output spectra are described by (5) and (6). For the other extreme (staggered clocks), each successive

driving the inverting

with gain thereafter.

to

The transient resulting from C7R charging

charge sample on CI because will only the if the trapped ampli-

V. does
of

trap

T/H

samples

the held 13(c) the

output where

of

the previous

stage.

This It is

charge tude

periods

remain

sampling spectrum easily

of a held of Fig. that

output

results in TSH = 1 and the output pb is the passband edge. envelope noise the held provides effective with larger

Vn
is

returns

to zero

at the instant

111 opens. ~1 opens

The (at

additional time tl)

charge

trapped

on CI at the instant

shown

sin x/x

anti-alias

filtering

of the folded

(increasing

f$/Pb

ratios) so that resampling As

increases the in-band noise. the output approximated

T/H output minimally staggered T/Hs are cascaded,

noise of the previous stages

AqcI(tl) with Avcl(tl)

= qCR = CR Un(tl)

noise of a given stage (in the hold phase) is closely by the resampled wide-band noise of the stage the present

= ~

Vn(tl).

added to the sampled

T/H

sampling. In summary, since most SCNS use staggered

is

clocks, TSH will

be taken to be unity for the rest of the paper, and the effects of resampling of noise will be ignored.

With S2 closed and SI open, qc~ is bled off to ground, and Vm with unity gain and a dc offset equal to the voltage stored on CI. When Sz opens and S1 closes, the inverting terminal is initially pulled to ground, and the op-amp then recharges CR to Pn, repeating the above procedure. Note
P. tracks
that the only time charge is trapped

VI. SCN INTEGRATORS AND CHARGE ACCUMULATORS In this section, a model of an SCN integrator is developed. The T/H results will then be combined with these results in Section VII where a simple SCN is analyzed.

on CI is when S1 opens, offset (

resulting in an accumulated
Vnet .

output

Vnet) of
(7)

y ~
i=1

V.(iT)

FISCHER: NOISE SOURCES AND SWITCHED CAPACITOR

FILTERS

747

+0
Fig. 15. SCN inte~ator noise model for CR/CI <<1 and Requ = Fig. 16. HPN noise test chcuit.

l/cRf~

where Jn(iT) is the noise voltage at the instant S1 opens, T is the complete SI, S2 period, and N is the number of complete clock periods since Cl was discharged. The output when Sl is closed is ~o.

R3

C4

R,

()
,+%

C*

Vn(t) vne*(N 1). + -

(8)

When S2 is closed, V.= Vn(t) + Vne(fv)

I
1~ c~
NON-INVERTING INTEGRATOR

(9)

Fig. 17. Noise simulation model for HPN SCN.

where the index of Vnet has been incremented because additional charge has been trapped on C[ when Sl opened. If there are N sample periods per clearing of CI (as in some
low-loss integration techniques), and Vn is Gaussian white

VII. NOISE MODEL AND CALCULATIONS

FOR SCNS

noise, standard words, larger

probability

theory

indicates

that

Vnet(N)

will have a

deviation {N the noise power

times larger than (CR /Cz) Vn. In other of Vnet after ~ cycles will be Al times noise power stored on CI in any

This section will develop a model suitable for SPICE noise simulations of switched capacitor networks. Experimental results in support of these results will also be given.

than

the incremental

A, SCN Topology Model

To simulate the capacitor switching noise (kT/C), replace the switched capacitors of the HPN in Fig. 16 with resistors of value

given sample cycle.

B. SCN Filter Approximations

A significant simplification can be made for those filter stages with pole frequencies much less than the sample rate. These filters usually have CR /CI <<1. Under these conditions, (7) shows that the individual contributions to Vnet are less than the referred to input noise Vn by CR /Cz. In addition, the number of terms in (7) that significantly contribute to the integrators output noise is limited by dc discharge paths from the amplifier output to its input. These discharge paths are used to provide Q damping and dc stability of the filter. In most practical SCNS, these conditions result in the input referred noise of the op-amp ( Vn) being much larger than Vnet, so the J&t term will be dropped. To model charge transfer uncertainty and the foldover effects of the wide-band noise of the switches associated with CR, an equivalent resistance of value [9]

R.~
Cf,
as per Fig. 17. Next replace the op-amps with the blocks shown in Fig. 17. The switched output (S.) is connected to all the switched capacitors connected to the output of the op-amps. The continuous output (CO) is connected to the paths that are not switched. To model the noninverting integrator of Fig. 16, use a voltage-controlled voltage source of gain = -1 to provide the polarity reversal (as in Fig. 17). The filter topology has now been modeled; next the op-amp model will be developed.

B. Op-Amp Model
To model the op-amp of Fig, 18 in an SCN environment, the sampled data effects on the output noise of the op-amp must be considered. This noise consists of two parts: wideband (flat, white noise), and l/f (flicker) noise. The voltage sources Vl,f and Vnl are used to model the unsampled l/f and flat-band noise, respectively. Since the folciover effects are added to the unsampled output noise, the noise sources kVnll and Vfll have been placed in series with the continuous output to yield SO. Referring to Appendix C to determine the op-amp BWn, the foldover factor (k) is derived by simplifying (6) under the constraints that .f<<.f., so the sin x/x term

cRf,

is used in the continuous-time equivalent circuit of Fig. 15. These approximations hold when the fflter passband is much less than the sample rate. For the noise simulations of the next section, the SCN integrator will be replaced by its continuoustime counterpart. The output noise of a particular SCN will be worked out in detail in Section VII to demonstrate how to use these results.

748

IN

v,/f

+ F
N,
1/f

IEEE JOURNAL

OF SOLID-STATE

CIRCUITS, VOL. SC-17, NO. 4, AUGUST

1982

co
IN co

so

so

F1l

kV

I?

v~,,

(f)

= ~

v,,,

= J-=

? USED TO MODEL R ~f,. RF44B #

IvN,,/rms

Ivfijm

BUT UNCORRELATED,
,oLDOv,R F/as,

co So

OUTPUT TO NON-SAMPLED PATHS OUTPUT TO SAMPLEO PATHS

Fig. 18. Frequency domain SCN op-amp noise simulation model.

Fig. 19. Complete SPICE op-amp noise model. The boxed circuit is the I/f source, md the resistors are uncorrelated white noise sources.

approaches unity and 7SH = 1. Then separate the result into terms related to the CO and SO outputs, respectively:

voltage as seen at the drain is

Vout(f) ~ % +v.

2BWn 1 [1 f.

n=is=%
To decouple the noise sources so that 1/f and wide-band noise effects can be studied separately, select the bias current and device W/L ratio to assure that the gm term contributes
much With less noise the device than the input-referred parameters selected, op-amp adjust flat-band noise.

with the bracketed term representing the foldover effects represented by lcVnll . Similarly accounting for the l/~ contribution (Appendix C), Q $ + (d).

KF to

match the

nl/f

out(f)

measured

op-amp input-referred

1/f noise. into the noninvertat the

Because the flicker

noise is to be injected

Because programs such as SPICE work with noise voltages rather than powers, the noise voltage gain through sampling is obtained by taking the square root of the bracketed terms to yield
k. i

ing terminal

of a high gain op-amp,

the dc bias voltage

drain must be blocked.

To avoid loading

the drain with the dc

-1 f,

2BWn

blocking circuit (l?llc, CBLK), a voltage-controlled voltage source is used as an ideal unity gain buffer, as in Fig. 19. The next subsection will show how to model the sampled I/f noise.

and the 1/f term of@. An alternate method of calculating k is as follows. First, calculate the total output noise power (PT) of the amplifier using SPICE or other means. The result is doubled to account for positive and negative sidebands contributing power, and is divided by the sample rate to yield an average power density. Putting this procedure into mathematics,

D. Flat-Band and Foldover Model

The flat-band, folded flat-band, and folded 1/f (Appendix C) can be modeled as white noise sources, which are easily simulated by resistors in SPICE. To model the flat-band noise, the resistor value is selected to match the measured input referred flat-band noise. To satisfy the nodal requirements of SPICE, two resistors (Rnl~ and I?nlB of Fig. 19), eac~ twice the value of the inputreferred noise resistance, are paralleled so that all nodes have at least two components connected and a dc path to ground. The source Vnl injects the unsampled flat-band noise into the op-amp, with Vnll modeling the wide-band foldover effects and Vjll the 1/f folded noise. Separate resistor sets are used to drive Vnl, Vnll, and Vll so that the sources will be mutually
uncorrelated. The appropriate folded I/f resistor values are

where T/FBis the presampled op-amp flat-band power density, The next two subsections will outline in detail how to use SPICE to model the op-amp noise.

C. Flicker Noise Model

Because SPICE lacks the option of user-defined functional expressions, a device model must be used to simulate the unsampled 1/f noise component. To conveniently generate the flicker noise, the boxed circuit of Fig. 19 is used. The device is diode-connected for ease of biasing to a predetermined drain current by ~&. To assure that the device is in the saturation region, choose Ml as an enhancement-type device. The noise

E.

Model Limitations

This model is accurate for f << f~, so the sin x/x frequency shaping will not appreciably affect the in-band noise. This

FISCHER: NOISE SOURCES AND SWITCHED CAPACITOR

FILTERS

749

TABLE III
MONOLITHIC FILTER RESULTS. NOISEVOLTAGES ARE rms

-410

VALUES INTEGRATED OVERTHEFREQUENCY RANGE FROM1 Hz TO 10 kHz. I/f

4.V
4..

FILTER
HPN
J, = 8kH,

FLAT CAPACITOR :.:;

6.5.. 3.9 25UV 9.01 4.1.?
45.3

dihc

MEASURED dBrnc
-6

12 6.81N 531N

-6.5

HPN
/s+4kHz

-11
6.7

-8
8.5

LPF
fs-128kHz

12.
12UV

LPF reduced caps.

25U

100V*

11 o

13

15

LPF reduced caps.& inputdcv.

I
o

I
2K

I
4K f IN HZ

6K

8K

10K

27UV 25UV

100UV4

11OW

13

16

O dBrn (24.5.v) CM = 0.6PF

1 Cw = 0.3rJF 1 C.m = L7pF 4 C.w = o.3pF

Fig. 20. Measured versus calculated 1.7 pF unit capacitor LPF noise response.

-80

constraint is not severe since most SCNS are designed with the sample rate much greater than the passband frequency. Further assumptio~s made to simplify the calculations are that the hold duty cycle of the switched paths (7sH) is unity and that the filter is run at a single sample rate. If stages of differing sample rates are cascaded, the sampled noise spectra of previous stages must be analyzed to account for significant foldover contributions to the stage under consideration.

90 -400 -420

L
0

I 8
, i

1 10
I

2 ,
1 ,

4
1 i

6
1

I
o

46

32
f IN KHz

40

64

80

F. Gdcutation and Measurement Results

Using models like the above, SPICE simulations were used to calculate the output noise for 60 Hz F@N filters sampled at 8 and 64 kHz, and several 128 kHz sample rate fifth-order elliptic low-pass filters with a cutoff frequency of 3.4 kHz [8]. The calculated C-message weighted (dBrnc) noise results are summarized in Table III, and are compared to results obtained from monolithic realizations of these circuits. The op-amp noise was measured to be 50 nV/ @ at 1 kHz, with a flat-band noise component of 35 nV/ @ and a noise bandwidth of 2 MHz. Fig. 16 shows an actual HPN (sampling at 8 kHz) switched capacitor filter [8], and its complete noise model is shown in Fig, 17, Calculating k for the filter with op-amps having a noise bandwidth (see Appendix C) of 2 MHz,

Fig. 21. Measured noise response of the 1.7 pF unit capacitor LPF for fs =128MZ (top trace) and for f$ = 1024 kHz (bottom trace). All analyzer settings are the same for both traces except that the lower trace stop frequency (80 kHz) is eight times that of the top trace (10 kHz).

in sample rate. As a test, the noise responses for ~~ = 128 and 1024 kHz of this LPF are compared in Fig. 21. All the analyzer settings are the same for both traces, with the exception that the frequency sweep was changed from 10 ,Wz up to 80 kHz [for ~~ = 1024 kHz) so the re~ative densities can be directly compared. The approximate 10 dB drop in the passband noise density clearly indicates that 1/f noise is definitely not the dominant noise source of these filters. VIII. CONCLUSION The results presented show that the topology of an SCN determines the significance of the relative contributions of amplifier flat-band and 1/f, and capacitor switching noise to the overall filter noise, For the test chip filter designs in our process, the dominant noise source was found to be capacitor switching noise, followed by op-amp wide-band noise, and lastly, op-amp 1/f noise. A noise model and a simulation technique have been develcrped to allow existing simulation programs (such as SPICE) to be used in accurate SCN noise calculations.
APPENDIX A

= FT5=-=223
Referring to Table V indicates that a = 2.1 for this sample rate. The dominant noise term of the HPN is the capacitor switching noise. If the falter is followed by an 8 kHz S/H, the wideband noise of the second amplifier will be boosted by the foldover factor (22.3) to become the dominant contributor. Similarly, the dominant noise source in the LPFs is the capacitor switching noise. The measured and predicted noise spectra are compared in Fig. 20 for the 1.7 pF minimum capacitor size LPF. If wideband noise sources are the dominant terms of the LPF, the noise density should drop by 9 dB for a factor af 8 increase

Transmission Gate Noise Mod@

The transmission gate simply m~dulates the input signal {x(t)} a pulse train, as shown in Fig. 22. The Fourier by

750

IEEE JOURNAL

OF SOLID-STATE

CIRCUITS, VOL. SC-17, NO. 4, AUGUST

1982

x(t)

~_y(+)

(+) ()

--P rl.--.r
x

X(t) =x(t) f(t)

J Y(t) T [x,] L~lzll~J TRANSFER FUNCTION [Ynl L-__J RECONSTRUCTION FILTER L___J SAMPLING OPERATION

Y(t)

TT

Fig. 22. Transmission gate model. TABLE IV TRANSMISS1ON NOISE RESPONSEAS A GATE CYCLE AND B WJf,
)UTY CYCLE T
100
1.00

Fig. 23. General topology of a sampled data system.

x(t)

FUNCTIONOF DUTY
%++f,~po

[h]

/ vowv.
B Wn as a function

INPuT

SAMPLER

of ~

SAMPLED DATA OUTPUT

J.
1 1Y2

Fig. 24. Ideal sampling switch operation.

10 I 00 094
079 0.59 0.49 034 019 0.090 0.040 0 I

1.00
0.95 0 so O 60 0.50 0.35 020 010 0049 0

.00

I .00 091 071 054 0.45 0.2s 011 0029 0007 0

1,00

095 0 so O 60 0,50 035 020 010 0.05 0

0.93 0 7s 059 0.4s 033 018 0 0S6 0.029 0

0.90 064 037 0.25 012 0.04 0.010 0003 0 constructs stitute the impulse held train output into a series of levels which conh T

Fig. 25. Ideal S/H impulse response.

the familiar

response.

Calculation of {x.} Spectrum

Taking the Fourier transform of {xn} [10], we find
X(ej2fT) =+ ~
kco

(fdd.v.r

eiTccts absent)

[1

27rf+*

(Bl)

series of the modulating

(-

pulse is

f(t)= T+; i=~ i -g

~)i i 27rt sin (i7rr)cos

()
T

where Xa is the Fourier transform of x(t), which is periodic about multiples of the sampling frequency. (Al)

Calculation of the S/H Frequency Response

Referring to the timing of Fig. 25, we can determine limits of integration: the

where ~ is the duty cycle and T is the sample period. Working out the details for a white noise input of bandwidth BWn,

%tlt(.f)=n.

[21F) H+in(
T +j ~ i+] or

A2
where

.Jf=_m

Jt=o

where h = BWn/j_~ (number of positive sidebands falling in band). Since (A2) is not in a convenient form for h, qout(~) has been tabulated in Table IV for various BWn /f~ and ~. An important approximation can be made for h >10:

= he-j2rf(k/2) sine (flz)

sine (x) =

sin (rrx) ~x

B W.
f,

%Nlt(f) w?? =

>10,

(A3)

h = (hold duty cycle) (sample period)

while at the other extreme, h = O (sin terms are not present), %ut(f) = ~% for 2BWn <f8. (A4) Substitution yields

APPENDIx B
A Sample-and-Hold Noise Model

G(2nf ) =rT sine r

When

A general topology for a discrete-time data system is shown in Fig. 23. The sampling function is shown in Fig. 24. The S/H operation consists of ideal sampling, followed by filtering with H(z) = 1, feeding into a hold operation. The hold re-

[01
K f.
noise,

e-iznf(hlz)

(B2)
the quantity of

dealing

with

power

is usually

interest;

calculating

the output

spectrum,

]Y(2flf)12 = ]X(ej2nfT) G(27rf)[2 = ]X(ej2TfT)[2 ]G(2nf)[2.

FISCHER:

NOISE SOURCES AND SWITCHED CAPACITOR FILTERS

OUTPUT NOISE DENSITY

751

Calculating lX(ei2zfT)12,

I I

For the case at hand, x(f) is white noise; hence, the aliases are uncorrelated, with the result that power, rather than voltage, terms are added. Using (2),
Fig, 26. Voltagefoltower
lX(e~2fT)12 +

output noise spectrum with test configuration.

r?. ~
1

()
for f,> r ~ f. .

for f, G 2BWn

(B3a)
h

l?n~

2BWn,

(B3b) ~f - fs (B4)
Fig. 27. Fundamental
-fof fs

and for the S/H,

]W)12 [01
r2T2
sinc2 Combining (B3) and (4) yields Jy(fJ2=2w%2sinc2 and Iy(f)l =?7n72 sinc2

and fust two sidebands of a sampled I/f noise source.

[(;)1s<2Bwn
be taken noise

to f.. The total op-amp output noise is calculated by adding up the power in each frequency band of width f~, The equivalent noise bandwidth (BWn) of an op-amp will
to be that bandwidth as the op-amp, required but with to contain a uniform the same spectral power

(B5a)

density qn.

[()1
I-~

f.>

2BWn.

(B5b)

I/f Foldover Effects Idealized sampling of a 1/f noise source produces the spectrum of Fig. 27 where only the first sidebands have been shown for simplicity, In the analysis that follows, it is assumed that the noise spectral density follows an ampl~ude envelope of the form A/f, and that the sidebands are mutually uncorrelated so that power, rather that voltage, is summed. Evaluating the foldover effects in the baseband from dc to fs,

The usual application of sampled data systems is the processing of information in the passband (f< f~/2); accordingly, the inband signal-to-noise ratio is of interest. Since signal-to-noise is given by s _ Iinput (2nf)12 [G(2nf)]2 X ]X(ej2TfT)[2 lG(27rf)12 =

Iinput (2mf)12
]X(ei2nfT)]2

it is independent of the reconstruction filter in the passband for white noise input, but does depend on the undersampling of noise. APPENDIX C

(cl)
where N is the number of sidebands folding into the baseband. To understand the foldover effects described in (Cl), let us assume that the voltage-follower op-amp attenuates the l/f noise for frequencies above 20 MHz, A = 1000 for computation ease, and that the frequency range of interest is the familiar telephone voiceband from 300 to 4 kHz. Fig. 28 compares the presampled and postsampled spectra for several common clock rates. An important aspect of Fig. 28 is that all the process dependence is lumped into the constant A, so these curves are simply multiplied by the same scale factor to fit any process. Evaluating just the foldover terms of (Cl) (and calling that sum a) indicates that the 1/f foldover effects can be closely approximated by adding a constant to the presampled spectrum. Also note that for typical sample rates of 64 kHz or higher, the foldover contribution is less than 20 percent of the 1/f baseband density at 1 kHz. A simplifying approximation is to neglect the 1/f foldover term entirely for a sample rate of 100 kHz or higher. For lower sample rates,

Foldover Effects and Equivalent Noise Bandwidth

This analysis will treat the op-amp 1/f and wide-band noise separately to clearly show the different effects of undersampling of these noise sources and to point out useful approximations.

Wide-BandNoise
Fig. 26 depicts a characteristic voltage-follower output noise spectrum (of the wide-band component only) that has been divided into frequency bands of width f,. Settling time constraints of SCNS usually restrict f~ to be much less than the op-amp unity gain frequency. Additionally, the switch bandwidths are usually set much wider than that of the op-amp. With the wide-bandwidth switches, effectively all the output noise of the op-amp is folded back in to the baseband from dc

752

IEEE JOURNAL

OF SOJ.ID-STATE CIRCUITS, VOL. SC-17, NO. 4, AUGUST

1982

7,,,

400

iK f IN Hz

4OK

Fig. 28. Pre- and postsampled l/~ spectra for several sample rates. TABLE V 1/y
FOLDOVER

FOI these curves,

V~/f =A/f with A = 1000.

FACTOR IN UNITS OF HZ-l

(Jz)
300 1000 2000 3000 3500 4000

.
= 8kHz 16kHz 32kHz 64Hz 12SkHz 256kHz

f,

2.1 2.1 2.1 2.2 2.2 2.2

0.96 0.96 0.97 0.97 0.97 0.97

0.44 0.44 0.44 0.44 0.44 0.44

0.20 0.20 0.20 0.20 0.20 0.20

0.09 0.09 0.09 0.09 0.09 0.09

0.04 0.04 0,04 0.04 0.04 0.04

place a white noise source in series with the sampled op-amp output (Fig. 16) of magnitude cd as summarized in Table V. ACKNOWLEDGMENT The author is grateful to P. E. Fleischer, A. Ganesan, D. G. Marsh, and K. R. Laker for their helpful discussions and ideas. Special thanks go to A. A. Schwarz for the excellent layout and laboratory work performed. REFERENCES
[1] C.-A. Gobet and A. Knob, Noise analysis of switched capacitor networks, in ISCAS Proc., Apr. 1981, pp. 856-859. [2] B. Furrer and W. Guggenbuhl, Noise analysis of sampled-data circuits, in ISCASF%oc., Apr. 1981, pp. 860-863. [3] M. L. Llou and Y-L. Kuo, Exact analysis of switched capacitor circuits with arbitrary inputs, IEEE Trans. Circuits Syst., vol. CAS-26, pp. 213-223, API. 1979. [4] A. B. Carlson, Communication Systems. New York: McGrawHill, 1968.

[5] R. A. Gable and R. A. Roberts, Signals and Linear Systems. New York: Wiley, 1973. rfil G. M, Jacobs, Practical design considerations for MOS switched capacitor ladder ffiters, Electron. Res. Lab., Univ. California, Berkeley, Memo. UCB/ERL M77/69, Nov. 1977. [7] D. J. Allstot, MOS switched capacitor ladder filters, Ph.D. dissertation, Electron. Res. Lab., Univ. California, Berkeley, Memo. UCB/ERL M79/30, May 1979. [8] P. E. Fleischer and K. R. Laker, A family of active switched capacitor biquad building blocks, Bell Syst. Tech. J., vol. 58, pp. 2235-2269, Dec. 1979. [9] C.-A. Gobet and A. Knob, Noise generated in switched capacitor networks, IEE Electron. Lett., vol. 16, pp. 734-735, Sept. 1980. [10] A. Peled and B. Liu, Digital Signal Processing. New York: Wiley, 1976. [11] P. E. Fleischer, A. Ganesan, and K. R. Laker, Effects of finite op amp gain and band-width on switched capacitor falters, presented at the IEEE Int. Symp. Circuits and Syst., Apr. 1981.
L-,

Jonathan H. Fischer received the B.S. and M.S. degrees in electrical engineering from the University of Colorado, Boulder, and the University of California, Berkeley, in 1978 and 1979, respectively. In 1978 he joined Bell Laboratories, Holmdel, NJ, as a member of the Technical Staff. Since then he has been involved in various aspects of charge-redistribution codec and filter work.