MISMATCH-SHAPING SERIAL DIGITAL-TO-ANALOG CONVERTER

Jesper Steensgaard

Un-Ku Moon and Gabor C. Temes

The Technical University of Denmark

Department of Information Technology
DK2800, Lyngby, Denmark

Oregon State University,

Department of Electrical Engineering,
ECE-220, Corvallis, OR 97331-3211

x n k

ABSTRACT
A simple but accurate pseudo-passive mismatch-shaping D/A converter is described. A digital state machine is used to control
the switching sequence of a symmetric two-capacitor network that
performs the D/A conversion. The error caused by capacitor mismatch is uncorrelated with the input signal and has only little power
in the signal band. The system has been simulated assuming a
0.1% capacitor mismatch, and the achieved SNDR performance
was 100 dB for an oversampling ratio of 7.

+
V

ref

(n 1) (n 2)

(n 3)

(n N )

reset

reset

( )

y n

MSB

( 1) v(n 2) v(n 3)

v n

reset

High-performance digital-to-analog converters (DACs) are used

for a variety of portable applications, e.g. in audio equipment. The
signal-to-noise-and-distortion ratio (SNDR) is often required to be
as large as 90100 dB, a level of performance which cannot be
obtained by brute-force implementation of DACs that rely on accurate matching of electrical parameters.
Error-shaping DACs have become popular because they facilitate linear D/A conversion despite the mismatch of electrical
parameters [13]. In this way, expensive laser trimming and elaborate background-calibration techniques can be avoided.
Portable equipment requires the power consumption be low,
and pseudo-passive DACs are potentially well suited for such applications. A pseudo-passive DAC containing only two capacitors,
a reference voltage source, and a few switches (Fig. 1) was described by Suarez et al. [4]. It functions by repeatedly charging C1
to Vref or 0 depending on the incoming bits x(n k) and sharing the
charge between C1 and C2 . It requires N clock cycles to convert
an N-bit digital word x(n) into an analog voltage y (n). A major disadvantage of this circuit is that its linearity is limited by the
matching of the two capacitors, which is typically no better than
10 to 11-bit accurate. Several recent papers [5, 6] discuss methods
for reducing this nonlinearity, but the algorithms described in them
require very complex logic and introduce new practical problems
associated with the precise addition of the analog outputs. This paper will discuss simpler techniques by which the DAC errors can
be reduced by error cancellation and/or mismatch shaping [7].

read

Clock Cycle

( ; 1)

( )

y n

y n

Figure 1: Basic two-capacitor DAC proposed by Suarez et al.

2 . Each D/A conversion
v (n 0) = 0 for all n.

is initiated by a reset phase reset , thus

Just before the k-th clock phase 2 , capacitor C2 will have

the voltage v (n k 1), and capacitor C1 will have the voltage
x(n k )Vref . During clock phase 2 , charge sharing yields the
following relation

v (n k )

C1
x( n
1 + C2

k )V

ref

C2
v (n
1 + C2

; 1)

(1)

By introducing the mismatch parameter

11 ; C12
11 + C12

(2)

eqn. (1) can be brought to the form

v (n k )

=
=

1+

2

x(n k )V

ref

x(n k )V

ref +

+ v (n
2

2. DAC ERROR ANALYSIS

x(n

k )V

v (n k

; 1)

ref ; v (n

; 1)]

; 1)
(3)

The first term in eqn. (3) represents the ideal operation

ideal (n) = videal (n

The research of U. Moon and G. Temes was supported by the NSF

Center for Design of Analog-Digital Integrated Circuits (CDADIC).

LSB

)
0

1. INTRODUCTION

The following analysis refers to the DAC shown in Fig. 1. Define

v (n k ) as the voltage across C1 and C2 in the k -th clock phase

read

x n k

x n k

N)

ref

N
X
k

=1

x(n k )2

; ;1 = Vref x(n)
N

(4)

The second term represents the error caused by capacitor mismatch. For all practical purposes, it is sufficient to consider only
the first-order errors1 , in which case
v (n k 1) can be approxP ;1
j ;k
imated by
Vref kj =1
x(n j )2
. This leads to the following
evaluation of the error signal

e(n)

=
=

y (n)

ref

ideal (n)

N
X
k

"

x(n k )

=1

;1
X

=1

;
j )2

x(n

#
k

; ;1(5)
N

Notice that the main sum in eqn. (5) is a function of x(n) only, i.e.,
contains the harmonics of x(n). The total harmonic distortion (THD) can, in principle, be calculated analytically except for
the unknown parameter
, to which the THD is proportional. Because an analytical expression will provide hardly any valuable insight, the THD performance was evaluated on the basis of simulations. It was found that, for a full-scale sinusoid input, the THD is
approximately ( 5 + 20 log10
) dB. Unfortunately, even when
using the best known layout techniques,
cannot be made much
smaller than 0.1%, corresponding to only about 11-bit linearity.
e(n)

jj

3.1. Error-Canceling DAC

8
; 8

x n k

3. MISMATCH-SHAPING DAC
The DACs signal-band performance can be improved dramatically by using mismatch shaping. To obtain this behavior, the error
signal e(n) must be made uncorrelated with x(n) (i.e., e(n) must
be a pseudo-noise signal), and its spectral power density must be
suppressed in the signal band.
In most mismatch-shaping DACs the error signal is controlled
by interchanging nominally identical components. For the twocapacitor DAC, the only degree of freedom is the choice of which
capacitor (C1 or C2 ) is charged in clock phases 1 . A new choice
can be made in every clock cycle. In the following, t(n k) = 1
will represent the condition that C1 is charged in the k-th clock
cycle of the n-th sample, whereas t(n k) = 1 will represent
that C2 is charged instead.
By revisiting eqns. (15), it is found that the error signal can
be calculated from

X
N

e(n)

ref

V

V

x(n k )

t(n k )b(n k )2

; ;1
N

k=1
ref b(n)

(6)

where
b(n k )

;1
X
k

=1

x(n j )2

(7)

Notice that e(n) is proportional to

b(n) =

N
X
k

=1

t(n k )b(n k )2

From eqn. (7) it can be observed that the coefficients b(n j ) j

1 2 3 : : : N are functions of the digital word x(n) only. Eqn.
(6) shows that, for a fixed x(n), the polarity of b
(n) can be alternated by inverting t(n k) k. Hence, if x(n) is D/A converted
twice using opposite but otherwise arbitrary values for t(n k) in
the two conversions, the two generated analog voltages ya (n) and
yb (n) will include errors ea (n) and eb (n) of the same magnitude,
but of opposite polarity. By adding or averaging ya (n) and yb (n),
the result will be a first-order error-free D/A conversion.
Active circuitry will be required to add ya (n) and yb (n), but
scaled averaging can be implemented passively as shown in Fig. 2.
For simplicity, t(n k) = 1 k in the first conversion a , and
consequently t(n k) = 1 k in the second conversion b . It
can be shown that, despite mismatch of the nominally identical
averaging capacitors Ca and Cb , only second- and higher-order
mismatch errors
p p > 1 will affect y (n).

)
(

x n k

2
1

read

read

Ca

Cb

( )

y n

read
a
b
y

Figure 2: Error-canceling D/A converter that requires 2N clock

cycles for the D/A conversion of a N -bit digital word x(n).
The main disadvantage of the described error-canceling DAC
is that 2N clock cycles are required for the D/A conversion of an
N -bit digital word x(n). If bandwidth is a problem, x(n) can be
interpolated to (say) N=2 bits of resolution using a low oversampling ratio of (say) 10 [8]. Notice that, to simplify the reconstruction filter processing y (n), this degree of oversampling is generally
required anyway.
The performance of the proposed system, using  interpolation, is equivalent to that of a mismatch-shaping DAC, although
the filtered pseudo-noise error signal is caused by the interpolation
rather than capacitor mismatch.
3.2. Mismatch-Shaping Switching

; ;1
N

(8)

and hence that the error signals spectral composition can be controlled without knowing the actual capacitor mismatch
. In other
words, the task is to choose t(n k) such that the spectral composition of b
(n) is of the desired form. A very simple error-canceling
DAC will be discussed first.
1 In other words, the terms that contain a coefficient
may be neglected.

, where

p >

1,

The operation of the two-capacitor DAC can be made mismatchshaping by choosing t(n k) appropriately in each clock cycle. An
advantage of this approach, compared to the error-canceling approach, is that the two averaging capacitors Ca and Cb can be
omitted and that only N clock cycles are required for the conversion of an N -bit word x(n). The objective is to choose t(n k)
such that b
(n) from eqn. (8) has the desired spectral composition.
Eqn. (7) shows that the coefficients b(n k) are functions of
x(n) only, and that b(n k )
1. Eqn. (8) shows that the set of
values x(n)] that b
(n) can attain using all possible combinations

j

of t(n k) depends only, but strongly, on the coefficients b(n k),

i.e., on x(n)2 . Because b(n k)
1 and t(n k ) = 1, and
because the b(n k)t(n k) are scaled by 2k;N , it follows that the
values in x(n)] will be affected mainly by the most significant
bits of x(n).

j

x(n) = 0000
x(n) = 1111

x(n) = 0001
x(n) = 1110

x(n) = 0010
x(n) = 1101

0.5

0.5

0.5

0.5

0.5

0.5

0.5

0.5

b(n)

1
0
1

1
0

12

x(n) = 0100
x(n) = 1011

1
0

12

x(n) = 0101
x(n) = 1010

0.5

0.5

0.5

0.5

0.5

0.5

0.5

0.5

2f

;g

x(n) = 0000
x(n) = 1111

1
0

12

x(n) = 0110
x(n) = 1001

x(n) = 0011
x(n) = 1100

wired read-only memories (a gate array). The LSBs of t(n), i.e.,

t(n k ) for k
1 2 : : : N 4 are chosen of the same/opposite
polarity as that of the respective b(n k) coefficients, such that
PN ;4


t(n k )b(n k ) will have the same polarity as that of  (n).
k=1
This can be implemented by performing exclusive-NOR operations on x(n k) and the sign bit of 
(n) (which is assumed to
be logic 1 when 
(n) is positive).

1
0

12

1
0

12

x(n) = 0111
x(n) = 1000

1
0

12

b(n)

1
1
1

x(n) = 0001
x(n) = 1110

x(n) = 0100
x(n) = 1011

1
1
1

1
1

Fig. 3 shows the attainable values of b

(n) for all combinations
of x(n) and t(n k) when represented with 4-bit accuracy. Eqn. (8)
was approximated by a summation from k = N 3 to k = N , and
in eqn. (7) x(n) was approximated by xxxx10000 : : : . The horizontal axes represent the binary value of the 4 MSBs of t(n k).
The sets x(n)] are found by projecting these values onto the
vertical axes. Clearly, b
(n) is a highly nonlinear function of x(n)
and t(n k). The important point is, however, that the function is
entirely numerical and independent of the capacitor mismatch
.
Hence, b
(n) can be controlled by choosing the sequence t(n k )
according to the present value x(n) and the desired value 
(n) of
b(n).

LSBs

( )

x n

Bitwise Ex-NOR

;4

=1

x(n) = 0101
x(n) = 1010

x(n) = 0011
x(n) = 1100

1
1

x(n) = 0110
x(n) = 1001

1
1

Figure 3: b
(n) as a function of the 4 MSBs of x(n) and t(n)

(n)

1
1
1

x(n) = 0111
x(n) = 1000

12

t(n k )

Sign of

x(n) = 0010
x(n) = 1101

12

1
0

1
1

 (n)

1
1

Figure 5: Control characteristic of b
(n) as a function of  (n).
Fig. 5 shows the sign selectors control characteristic, i.e., the
(n), when the sign selector is implerelation between 
(n) and b
mented as shown in Fig. 4. Clearly, the controllability of b
(n) and
the control characteristics linearity depends on the digital input
word x(n); this reflects the composition of x(n)]. The controllability of b
(n) is, however, much better than that of a traditional
(n) can attain only two
single-bit  modulator (Fig 6) where b
values x(n) 1 and x(n) + 1, neither of which necessarily will
be close to 
(n).

x(n)

;4

b(n)

H (z )

(n)

MSBs

LSBs

(n)

MSBs

ROM
(Hard-Wired)

LSBs

( )

t n

MSBs

Discard

Figure 6:  modulator showing the equivalents of b

(n) and


 (n).
Fig. 7 shows a block diagram of the overall DAC system. The
modulator can be implemented using serial logic because it is
clocked only once for each new sample x(n).


Figure 4: Sign selector choosing t(n

a value close to 
(n).

k)

such that b
(n) will attain

P/S

Fig. 4 shows how the control system, which will be called a

sign selector, can be implemented. The four MSBs of t(n), i.e.,
t(n k ) for k
N
3 N
2 N
1 N are selected on the
basis of Fig. 3 as the code that yields the value of b
(n) which is
the closest to 
(n) truncated to 4-bit representation. This operation can, for example, be implemented by four small 4x4 bit hard-

2f ;

property is shared by single-bit  modulators, where x(n)]

is a set of only two values: x(n) 1 and x(n) + 1. For the two-capacitor
DAC, however, x(n)] may contain as many as 2N elements grouped in
pairs of equal value but opposite polarity.
2 This

Parallel-to-serial conversion

x(n)

Error
Estimator
Eqns. (7,8)

b(n)

()

H z

(n)

Sign
Selector
Fig. 4

Twoy(n)
Capacitor

t(n)

DAC

P/S
Vref

Figure 7: Mismatch-shaping two-capacitor DAC system.

The sign selector, shown in Fig. 4, can be simplified by reducing the ROMs resolution, but the corresponding control characteristic will typically be less linear. In the extreme case, the ROM

may be entirely omitted, and then the control characteristic will

resemble that of a single-bit delta-sigma modulator. The linearity
required of the control characteristic depends on how aggressively
the loop filter H (z ) is designed. If the control characteristic is very
nonlinear the maximum gain NTFmax of the noise transfer functions NTF(z ) = 1=(1 + H (z )) should be less than (say) 1.5 [1].
However, if the control characteristic is fairly linear, NTFmax may
(n) will be
be chosen larger, and the signal-band suppression of b
improved [8].
The loop filter H (z ) need not be of high order because the
Nyquist-band error power is low (in the order of 60 dB). Considering that the control characteristic is always quite nonlinear
for low-level inputs, i.e., when the MSBs of x(n) is 1000 or
0111, it makes good sense to use a simple (second-order) loop
filter because it always will be stable3 .

To avoid very complex reconstruction filters, most DACs will

convert somewhat oversampled signals. Even if the oversampling
ratio is as low as 7 to 10, 100 dB signal-band performance can be
obtained when using the mismatch-shaping DAC shown in Fig. 7.
It requires more digital circuitry, but even so, the overall power
consumption may be lower than that of the error-canceling DAC.
If the loop filter is of low order, the control characteristic may be
quite nonlinear and the sign selector can be made very simple. The
operation is based on a mathematical estimate of the capacitor mismatch error that is accurate to a first-order approximation. Other
higher-order errors not accounted for are essentially negligible.
Although not discussed in this paper, the two-capacitor DAC
can be made insensitive to charge-injection and clock-feedthrough
errors [9].

5. REFERENCES
0

50

50

100

100

10

10

(a) dB vs.

10

f =fs

[1] Steven R. Norsworthy, Richard Schreier, and Gabor C. Temes,

Eds., Delta-Sigma Data Converters: Theory, Design, and
Simulation, IEEE Press, 1996.

10

10

(b) dB vs.

10
f =fs

Figure 8: Simulated (FFT) performance of the DAC without (a)

and with (b) mismatch shaping using a second-order loop filter
H ( z ).
Fig. 8 shows the simulated performance when the relative capacitor mismatch was 0.1%. The peak-to-peak value of y (n) was
0.71Vref , which is defined as 0 dB. Fig. 8a shows the performance
of the basic DAC shown in Fig. 1. Harmonic spurs are observed;
the THD performance is approximately 71 dB. Fig. 8b shows
the performance of the mismatch-shaping DAC shown in Fig. 7,
for which NTF(z ) = 1 1:91z;1 + z ;2 and a small amount of
dither was added to 
(n) to prevent idle tones. Harmonic spurs
are not observed, and the mismatch-shaping property is observed.
The error signals signal-band power is shown as the dashed line;
the performance is better than 100 dB even for oversampling ratios
as low as 7.

4. CONCLUSIONS
The discussed pseudo-passive two-capacitor DAC structure is suitable for portable applications because it uses only little dc power.
The serial operation limits the bandwidth, but it is wide enough for
audio application, for example. Two techniques were proposed to
avoid the harmonic distortion caused by capacitor mismatch. The
error-canceling DAC, shown in Fig. 2, is simple to implement and
will yield a very good performance. The bandwidth-to-power ratio, however, is cut in half because each digital word is converted
twice.
3 Higher-order loop filters can designed and successfully used, but it is
advisable to stabilize the  modulator by hard-limiting the state variables
representing higher-order summations of b
(n), or by incorporating reset
operations when the system becomes unstable.

[2] Robert Adams, Khiem Nguyen, and Karl Sweetland, A

113dB SNR Oversampling DAC with Segmented NoiseShaped Scrambling, in Digest of Technical Papers for
the 1998 International Solid-State Circuits Conference, San
Fransisco, February 1998, IEEE Solid-State Circuits Society,
vol. 41, pp. 6262.
[3] Richard Schreier and B. Zhang, Noise-shaped multibit D/A
converter employing unit elements, Electronics Letters, vol.
31, no. 20, pp. 17121713, September 1995.
[4] R. Suarez, P. Gray, and D. Hodges, All MOSFET chargeredistribution analog-to-digital conversion techniquesPart
II, IEEE Journal of Solid-State Circuits, vol. SC10, no. 7,
pp. 379385, December 1975.
[5] L. Weyten and S. Audenaert, Two-capacitor DAC with compensative switching, Electronics Letters, vol. 31, no. 17, pp.
14351436, August 1995.
[6] P. Rombouts, L. Weyten, J. Raman, and S. Audenaert, Capacitor Mismatch compensation for quasi-passive switchedcapacitor DAC,
IEEE Transactions on Circuits and
SystemsI: Fundamental Theory and Applications, vol. 45,
no. 1, pp. 6871, January 1998.
[7] Jesper Steensgaard, Un-Ku Moon, and Gabor Temes,
Mismatch-Shaping Switching for Two-Capacitor DAC,
Electronics Letters, vol. 34, no. 17, pp. 16331634, August
1998.
[8] Richard Schreier,
Mismatch-Shaping Digital-to-Analog
Conversion, An Audio Engineering Society Preprint of paper [4529 (E-1)] presented at the 103rd convention, September
1997, New York.
[9] Jesper Steensgaard, Un-Ku Moon, and Gabor C. Temes,
Mismatch-Shaped Pseudo-Passive Two-Capacitor DAC, in
Conference Proceedings for the 1999 IEEE Alessandro Volta
Workshop on Low-Power Design (VOLTA99), held March 45, 1999 in Como, Italy, 1999.