AD676
REV. A
–13–
AC PERFORMANCE
AC parameters, which include S/(N+D), THD, etc., reflect the
AD676’s effect on the spectral content of the analog input sig-
nal. Figures 12 through 16 provide information on the AD676’s
ac performance under a variety of conditions.
As a general rule, averaging the results from several conversions
reduces the effects of noise, and therefore improves such param-
eters as S/(N+D). AD676 performance may be optimized by
operating the device at its maximum sample rate of 100 kSPS
and digitally filtering the resulting bit stream to the desired signal
bandwidth. This succeeds in distributing noise over a wider
frequency range, thus reducing the noise density in the fre-
quency band of interest. This subject is discussed in the follow-
ing section.
OVERSAMPLING AND NOISE FILTERING
The Nyquist rate for a converter is defined as one-half its sam-
pling rate. This is established by the Nyquist theorem, which re-
quires that a signal he sampled at a rate corresponding to at
least twice its highest frequency component of interest in order
to preserve the informational content. Oversampling is a conver-
sion technique in which the sampling frequency is more than
twice the frequency bandwidth of interest. In audio applications,
the AD676 can operate at a 2 3 F
S
oversampling rate, where
F
S
= 48 kHz.
In quantized systems, the informational content of the analog
input is represented in the frequency spectrum from dc to the
Nyquist rate of the converter. Within this same spectrum are
higher frequency noise and signal components. Antialias, or low
pass, filters are used at the input to the ADC to reduce these
noise and signal components so that their aliased components
do not corrupt the baseband spectrum. However, wideband
noise contributed by the AD676 will not be reduced by the
antialias filter. The AD676 quantization noise is evenly distrib-
uted from dc to the Nyquist rate, and this fact can be used to
minimize its overall affect.
The AD676 quantization noise effects can be reduced by
oversampling–sampling at a rate higher than that defined by the
Nyquist theorem. This spreads the noise energy over a band-
width wider than the frequency band of interest. By judicious
selection of a digital decimation filter, noise frequencies outside
the bandwidth of interest may be eliminated.
The process of analog to digital conversion inherently produces
noise, known as quantization noise. The magnitude of this noise
is a function of the resolution of the converter, and manifests it-
self as a limit to the theoretical signal-to-noise ratio achievable.
This limit is described by S/(N+D) = (6.02n + 1.76 + 10 log
F
S
/2F
A
) dB, where n is the resolution of the converter in bits, F
S
is the sampling frequency, and Fa is the signal bandwidth of in-
terest. For audio bandwidth applications, the AD676 is capable
of operating at a 2 3 oversample rate (96 kSPS), which typically
produces an improvement in S/(N+D) of 3 dB compared with
operating at the Nyquist conversion rate of 48 kSPS. Over-
sampling has another advantage as well; the demands on the
antialias filter are lessened. In summary, system performance is
optimized by running the AD676 at or near its maximum sam-
pling rate of 100 kHz and digitally filtering the resulting spec-
trum to eliminate undesired frequencies.
DC CODE UNCERTAINTY
Ideally, a fixed dc input should result in the same output code
for repetitive conversions. However, as a consequence of system
noise and circuit noise, for a given input voltage there is a range
of output codes which may occur. Figure 9 is a histogram of the
codes resulting from 1000 conversions of a typical input voltage
by the AD676 used with a 10 V reference.
2
10–1
DEVIATION FROM CORRECT CODE – LSBs
NUMBER OF CODE HITS
800
0
200
400
600
Figure 9. Distribution of Codes from 1000 Conversions,
Relative to the Correct Code
The standard deviation of this distribution is approximately 0.5
LSBs. If less uncertainty is desired, averaging multiple conver-
sions will narrow this distribution by the inverse of the square
root of the number of samples; i.e., the average of 4 conversions
would have a standard deviation of 0.25 LSBs.
