Message-ID: <4ED90E32.6080507@hifidelity.com>
Date: Fri, 02 Dec 2011 17:43:14 +0000
From: Steve Dove <dsp@hifidelity.com>
User-Agent: Mozilla Thunderbird 1.0.6 (Windows/20050716)
MIME-Version: 1.0
To:  rsgb_lf_group@blacksheep.org
References: <E1RW9bG-0001fW-00.rn3aus-mail-ru@f214.mail.ru> <E1RWVQQ-00015Y-K7@post.thorcom.com>
In-Reply-To: <E1RWVQQ-00015Y-K7@post.thorcom.com>
Subject: Re: LF: 16 bit vs 24 bit ADC?
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Sender: owner-rsgb_lf_group@blacksheep.org
Precedence: bulk
Reply-To: rsgb_lf_group@blacksheep.org
X-AOL-SCOLL-URL_COUNT: 0  
x-aol-sid: 3039ac1d60154ed90ef21182
X-AOL-IP: 195.171.43.25
X-AOL-SPF: domain : blacksheep.org SPF : none


Hi Bill,

Audio afficionados are high on opinion but notoriously 
hard-put to show statistically significant discrimination in 
true ABX or ABC scenarios; I wouldn't be using them as any 
sort of criterion.  That said, there are strong reasons for 
using 24-bit S/D audio convertors, some qualitatively based 
involving enhanced dynamic range, but mostly that it is quite 
difficult to find anything else nowadays!  (Except perhaps in 
the icky depths of soundcard-land.)

But '24 bits' ain't what it seems  -  even the *best* 
integrated audio convertors *well implemented* are 'only' 
capable of 120dB dynamic range measured honestly.  Even though 
the internal architecture/filters etc. may be 24 bits wide, 
it's in effect a 20 bit convertor with 4 'marketing bits'. 
With luck the noise is incoherent enough to use as free 
dither, but that isn't guaranteed.  Lesser parts, the lower 
cost ones more commonly found in consumer-world rather than 
pro-audio, tend to come out shy of 100dB, at best about the 
same as (non-existent) 'perfect' 16-bit convertors.

It is a long, long time since I've been able to see the lowest 
bit of a convertor toggle with applied signal (and turn into a 
pseudo-PWM!) - they're ALWAYS noise-bound nowadays.  The 
question turns into whether the device's noise is going to 
dominate, or yours (the application; usual).

So, in audio (or as used baseband in SDR) we're stuck with a 
mere 120dB dynamic range in a 20kHz bandwidth.  Which actually 
isn't shabby at all, giving 'pure analogue' a run for its money.

As for the way-more-serious ultra-fast convertors used in 
direct-SDR it is to be presumed that even the latest 
generation of 16-bit monsters are similarly noise-bound.

It's common practise (at least during development) to 
deliberately run the convertors at very low level and 
listen/measure amplified up to see what the noise floor 
sounds/looks like.  The better the implementation, the less 
coherent crap and the noisier the noise.  Cheaper parts are 
often riddled with internal squirglies, the good ones less so. 
  THIS - the nature of the noise-floor - could have a far 
greater bearing on useful dynamic range in a radio application 
than mere number of bits.

But you raise a very valuable question that transcends the 
'what is possible'.  It's how much dynamic range do we really 
NEED?


         73,

                 Steve




Bill de Carle wrote:
> We have seen a tendency to use ultra narrow bandwidths to coax out weak 
> signals not otherwise discernable.  This calls for enhanced frequency 
> stability at both ends, but that is not the only factor involved.  At 
> some stage we use an ADC to convert the analog signal to digital, 
> thereafter working only with numerical approximations of instantaneous 
> voltages: that's where information can be lost.
> 
> Traditionally, 16-bit ADC's have served well - they're readily available 
> and inexpensive.  At some point though, the signal we are looking for 
> becomes just too small to register on a 16-bit ADC.  It seems intuitive 
> that in a pristine noiseless environment, if a feeble sinewave has 
> insufficient amplitude to toggle the LSB of our ADC it might as well not 
> exist at all.  Clearly, if a signal produces no effect on a radio 
> receiver (where "receiver" includes the ADC looking at its output) - the 
> end result is the same as if that signal wasn't there.  No amount of 
> post-processing or extended sampling time is ever going to detect it.
> 
> One might argue that adding a little noise such that the sum of noise + 
> weak signal will then be enough to change the ADC output will solve the 
> problem.  Unfortunately things aren't that easy.  Let's remember that an 
> ADC *approximates* some true value; the quality of such an approximation 
> isn't entirely specified by the number of bits used to represent the 
> ADC's output, there are other factors, in particular the "correctness" 
> of the result.  For example, say an 8-bit ADC is supposed to resolve 7 
> amplitude bits plus a sign.  If the full-scale input is specified to be 
> 1 volt, the least significant bit represents 7.8125 millivolts.  
> Assuming all the other bits are zeros, for the LSB to be honest, it 
> should say "1" if the input voltage is greater than or equal to 3.390625 
> millivolts (i.e. half of 7.8125 mV), and "0" otherwise.  ADC 
> manufacturers will guarantee monotonicity, but that's much easier to 
> achieve than the absolute honesty we'd like.
> 
> We sometimes take many consecutive readings of a parameter then average 
> them to get a better idea of the true value.  If we take hundreds of 
> measurements known to be correct to one decimal place, we might have 
> some confidence in expressing the average to 2 decimal places.  It would 
> be justified if 1) the parameter being measured doesn't change 
> significantly over the sampling period (or it changes in a constrained 
> way e.g. in an FFT) and 2) the ADC is honest.  Any intrinsic error in 
> our ADC (say the one above has a comparator offset and reports 3.5000 
> millivolts as "0" instead of "1") cannot be "corrected" by long term 
> averaging.  Once information is lost it is unrecoverable.  Even with 
> "dithering" (adding some white noise to jostle the LSB), our long term 
> average will still be limited by the honesty of the ADC.
> 
> It has been said that weak signal reception is all about the 
> signal-to-noise ratio, not about the strength of the signal itself. I 
> think in today's digital world with limited dynamic range that old adage 
> needs to be rethought.  Remember that truncating a numerical 
> approximation to any number of bits less than infinity is equivalent to 
> adding noise.
> 
> Since manufacturers usually guarantee monotonicity, there is reason to 
> believe a 24-bit ADC would be more "honest" than a 16-bit ADC, i.e. will 
> give us numbers which are closer to the true underlying values.  Audio 
> afficionados sometimes claim they can hear the difference, saying a 
> 24-bit ADC sounds better to them, often using a lot of non-scientific 
> words to describe the improvement.  My question is this: In a practical 
> situation with low QRN/QRM, would going to a 24-bit ADC soundcard result 
> in an ability to detect a weak LF signal that would not show up on a 
> Spectrum Lab display computed from 16-bit values?
> 
> Bill VE2IQ
> 
> 
> 
> 
> 
> -----
> No virus found in this message.
> Checked by AVG - www.avg.com
> Version: 2012.0.1873 / Virus Database: 2102/4652 - Release Date: 12/02/11
> 
> 
>