All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

Suppose we bought four of the best stereo ADCs money could buy. Dan
Lavry mentioned his unit that has 127dB dynamic range, equal to more
than 21 bits of "useful" data. Okay, so suppose we fed a central clock
to four of these units, and fed all four of them with a single stereo
signal. Now we take the digital outputs of these boxes and record them
into a 24-bit, 8-channel recorder or DAW. Now suppose we took the
resulting data and averaged all four stereo recordings into one new
stereo recording. Intuitively I think this should result in a
theoretical increase in dynamic range of 6dB, for a total of 133. I'm
figuring an improvement of 3dB for each doubling the number of
converters. Is that right? In theory? So if we really wanted to,
couldn't we have as many bits of real resolution as we wanted, in a
single box, by paralleling massive numbers of ADCs and averaging them
in an onboard DSP process to output a single data stream? Is this
concept in use today? I seem to remember some company (dcs?) messing
with this at some point, but I don't really remember.

ulysses

All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

Suppose we bought four of the best stereo ADCs money could buy. Dan
Lavry mentioned his unit that has 127dB dynamic range, equal to more
than 21 bits of "useful" data. Okay, so suppose we fed a central clock
to four of these units, and fed all four of them with a single stereo
signal. Now we take the digital outputs of these boxes and record them
into a 24-bit, 8-channel recorder or DAW. Now suppose we took the
resulting data and averaged all four stereo recordings into one new
stereo recording. Intuitively I think this should result in a
theoretical increase in dynamic range of 6dB, for a total of 133. I'm
figuring an improvement of 3dB for each doubling the number of
converters. Is that right? In theory? So if we really wanted to,
couldn't we have as many bits of real resolution as we wanted, in a
single box, by paralleling massive numbers of ADCs and averaging them
in an onboard DSP process to output a single data stream? Is this
concept in use today? I seem to remember some company (dcs?) messing
with this at some point, but I don't really remember.

I think you're thinking of balanced a/d, where a stereo a/d is fed a mono
signal in one channel and the inverted mono signal to the other, then the
inverted signal is flipped back, and summed to the original. Apparently it
does improve S/N, but I believe it is already applied inside high-end a/d
converters.

What exactly would you do with the extra 6dB of dynamic range anyway? I
doubt there's a mic on the planet with better self-noise than -96dB. Even
if there was, you'd need about 60dB of gain to make any such subtleties
audible. I think the advantages of 24-bit sampling have more to do with
minimized rounding errors than dynamic range.

Justin Ulysses Morse wrote in message om...
All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

Yes, it's been done.

Suppose we bought four of the best stereo ADCs money could buy. Dan
Lavry mentioned his unit that has 127dB dynamic range, equal to more
than 21 bits of "useful" data.

Depends on how he defines "dynamic range". It could be spurious-free
dynamic range, full-scale minus noise floor, THD+N expressed in dB(A),
or something else. Based on what he's measuring, the number may or may
not improve when you add identical redundant ADCs.

Okay, so suppose we fed a central clock
to four of these units, and fed all four of them with a single stereo
signal. Now we take the digital outputs of these boxes and record them
into a 24-bit, 8-channel recorder or DAW. Now suppose we took the
resulting data and averaged all four stereo recordings into one new
stereo recording. Intuitively I think this should result in a
theoretical increase in dynamic range of 6dB, for a total of 133. I'm
figuring an improvement of 3dB for each doubling the number of
converters. Is that right? In theory?

It's right if you believe the "dynamic range" is dominated by
random noise, not by distortion products. Only the noise will
be reduced by averaging.

So if we really wanted to,
couldn't we have as many bits of real resolution as we wanted, in a
single box, by paralleling massive numbers of ADCs and averaging them
in an onboard DSP process to output a single data stream?

You might get some improvement, but it's ultimately going to be
limited by distortion terms (see above), clock jitter sidebands,
and leakage from clocks and power supplies.

Now an interesting thing to do is NOT to synchronize the ADC's,
as you assumed above. Instead, use a high-quality sample rate
converter on each ADC output prior to combining them. Then you
could get more error terms to average out, though I think
input-related distortion still wouldn't.

Is this
concept in use today? I seem to remember some company (dcs?) messing
with this at some point, but I don't really remember.

Daniel Weiss used 4 (I think) delta sigma ADC chips synchronized
together. Now that asynchronous SRC chips have gotten so good,
it's practical to use unsynchronized ADCs. I know of one company
that's presently considering that.

David L. Rick
Hach Company (the day job)
Seventh String Recording (sleepless nights)

Real humans are invited to ignore the header address and reply to:

davidDOTrickAThachDOTcom

Justin Ulysses Morse wrote in message om...
All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

Suppose we bought four of the best stereo ADCs money could buy. Dan
Lavry mentioned his unit that has 127dB dynamic range, equal to more
than 21 bits of "useful" data. Okay, so suppose we fed a central clock
to four of these units, and fed all four of them with a single stereo
signal. Now we take the digital outputs of these boxes and record them
into a 24-bit, 8-channel recorder or DAW. Now suppose we took the
resulting data and averaged all four stereo recordings into one new
stereo recording. Intuitively I think this should result in a
theoretical increase in dynamic range of 6dB, for a total of 133. I'm
figuring an improvement of 3dB for each doubling the number of
converters. Is that right? In theory? So if we really wanted to,
couldn't we have as many bits of real resolution as we wanted, in a
single box, by paralleling massive numbers of ADCs and averaging them
in an onboard DSP process to output a single data stream? Is this
concept in use today? I seem to remember some company (dcs?) messing
with this at some point, but I don't really remember.

ulysses


ulysses
I am not sure i understand the theoretical basis for your assertion
that doubling the number of converters would result in a 3db increase
in dynamic range. could you explain this further?
Rich

You're assuming that each ADC has only random errors, which may not be
true - especially if each ADC unit is identical. A slightly better setup
could use totally different units that are capable of the same output
format, have the same gain, etc. Even using different units may not give you
totally random errors, though, as I imagine it is possible that high-end
ADC's have similar systematic flaws.

Ryan

I am not sure i understand the theoretical basis for your assertion
that doubling the number of converters would result in a 3db increase
in dynamic range. could you explain this further?
Rich

If you sum two identical signals, you end up with a signal that is 6
dB louder. If you sum two noise-sources (which are not correlated but
have the same RMS-value), the RMS-value of the resulting noise will be
3 dB higher. So if you sum two noisy signals, the signal itself will
get 6 dB more, the noise only 3 dB. So you get 3 dB more dynamic.
Samuel

"Ryan Mitchley" wrote in
message m
You're assuming that each ADC has only random errors, which may not be
true - especially if each ADC unit is identical. A slightly better
setup could use totally different units that are capable of the same
output format, have the same gain, etc. Even using different units
may not give you totally random errors, though, as I imagine it is
possible that high-end ADC's have similar systematic flaws.

IME the comments about systematic errors adding are pretty factual.

As I've pointed out in other posts, the noise in modern chip converters is
dominated or at least significantly composed of noise (essentially dither)
that is systematically generated on the converter chip. I don't know to what
degree these chips tend to come up in synch if just naively paralleled as
far as power and clock signals go.

Justin Ulysses Morse wrote:
All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

You're a litttle bit too late. Lots of the standard converters do this
internally already. The 1965 Bell Labs book _Digital Signal Processing_
describes it and goes through the math.
--scott

--
"C'est un Nagra. C'est suisse, et tres, tres precis."

See, I knew it would be an interesting conversation. Thanks for all
the input.

ulysses


In article , Justin
Ulysses Morse wrote:

All this great discussion of digital audio theory, bits, sample rates,
dither, etc. got me thinking about something: Parallel ADC for
increased resolution.

Suppose we bought four of the best stereo ADCs money could buy. Dan
Lavry mentioned his unit that has 127dB dynamic range, equal to more
than 21 bits of "useful" data. Okay, so suppose we fed a central clock
to four of these units, and fed all four of them with a single stereo
signal. Now we take the digital outputs of these boxes and record them
into a 24-bit, 8-channel recorder or DAW. Now suppose we took the
resulting data and averaged all four stereo recordings into one new
stereo recording. Intuitively I think this should result in a
theoretical increase in dynamic range of 6dB, for a total of 133. I'm
figuring an improvement of 3dB for each doubling the number of
converters. Is that right? In theory? So if we really wanted to,
couldn't we have as many bits of real resolution as we wanted, in a
single box, by paralleling massive numbers of ADCs and averaging them
in an onboard DSP process to output a single data stream? Is this
concept in use today? I seem to remember some company (dcs?) messing
with this at some point, but I don't really remember.

ulysses