On Topic, but Just For Fun

All,

This is for real.

A 16-bit A/D manufacturer showed a spectral plot to indicate the
performance of their device.

The spectrum was from digitizing a single tone, plotted Magnitude(dB)
vs frequency (Hz). There was a fairly flat noise floor (grass if you
prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
frequency expected.  The peak was 96 dB above the noise floor.

What can you say about the device performance?

Dirk

Dirk Bell
DSP Consultant

"dbell" <bellda2005@cox.net> wrote in news:1158782000.048968.277230
@k70g2000cwa.googlegroups.com:

> All,
> 
> This is for real.
> 
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
> 
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
> 
> What can you say about the device performance?
> 
> Dirk
> 
> Dirk Bell
> DSP Consultant
> 
> 


It works quite well when the tone being sampled covers the full scale of 
the A/D range??


-- 
Scott
Reverse name to reply

dbell wrote:


> This is for real.
> 
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
> 
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
> 
> What can you say about the device performance?

Looks like the manufacturer has no clue.

1. When presenting a spectral plot, one should tell what are the units. 
The spectral plot without the units does not make any sense. What is 
96dB?  Is it in the dBc/Hz^1/2 or what?

Assuming -96dBc over the whole bandwidth means the total power of the 
trash equals to that of the signal :)))))

2. The good ADC should have the trash concentrated in the peaks at the 
itnermod products of F and Fsa.

Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

"dbell" <bellda2005@cox.net> wrote in news:1158782000.048968.277230
@k70g2000cwa.googlegroups.com:

> All,
> 
> This is for real.
> 
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
> 
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
> 
> What can you say about the device performance?
> 
> Dirk
> 
> Dirk Bell
> DSP Consultant
> 

A measurement from the peak of the carrier to the top of the next highest 
peak (and usually exluding some number of filter bins or bandwidth around 
the carrier) is referred to as SFDR:  Spur Free Dynamic Range.

Presumably, you have some other information to go with this plot, because 
the result is dependant on:
The actual magnitude of the input carrier. (Usually 1 to 6 dB below a full 
scale input)
The sampling rate.
The input carrier's frequency.
The cleanliness of the clock source.
The cleanliness of the carrier source.
If the spectral plot was the result of an FFT, you would want to know what 
window function was used.
How much averaging was performed on the output.
Whether dithering was used.

Vendors will often provide a plot of typical SFDR vs input frequency at a 
specified sampling rate and input amplitude.

Whether SFDR is a useful parameter depends upon your application.  I find 
it very useful when I am going to be looking at my sampled signal in the 
frequency domain, 


-- 
Ian Shef     805/F6      *    These are my personal opinions    
Raytheon Company         *    and not those of my employer.
PO Box 11337             * 
Tucson, AZ 85734-1337    * 

Vladimir Vassilevsky wrote:
> dbell wrote:
>
>
> > This is for real.
> >
> > A 16-bit A/D manufacturer showed a spectral plot to indicate the
> > performance of their device.
> >
> > The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> > vs frequency (Hz). There was a fairly flat noise floor (grass if you
> > prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> > frequency expected.  The peak was 96 dB above the noise floor.
> >
> > What can you say about the device performance?
>
> Looks like the manufacturer has no clue.
>


> 1. When presenting a spectral plot, one should tell what are the units.
> The spectral plot without the units does not make any sense. What is
> 96dB?  Is it in the dBc/Hz^1/2 or what?

96dB *is* a unit :)

It needs no other qualifier (for a raw power statement, anyway). What
it means is something else, of course ;)

In this case, it stands for the unspectacular statement that a true 16
bit device has 96dB of S/N ratio. What they are claiming, however,
would have to be proved - a real 16 bit enob in a 16 bit device, even
ignoring issues such as thermally generated noise (which we are not
privy to as we don't know the absolute scale used) :)


Cheers

PeteS



>
> Assuming -96dBc over the whole bandwidth means the total power of the
> trash equals to that of the signal :)))))
>
> 2. The good ADC should have the trash concentrated in the peaks at the
> itnermod products of F and Fsa.
>
> Vladimir Vassilevsky
> 
> DSP and Mixed Signal Design Consultant
> 
> http://www.abvolt.com

"dbell" <bellda2005@cox.net> writes:

> All,
>
> This is for real.
>
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
>
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
>
> What can you say about the device performance?

Without knowing the "resolution bandwidth" - nothing.
-- 
%  Randy Yates                  % "She's sweet on Wagner-I think she'd die for Beethoven.
%% Fuquay-Varina, NC            %  She love the way Puccini lays down a tune, and
%%% 919-577-9882                %  Verdi's always creepin' from her room." 
%%%% <yates@ieee.org>           % "Rockaria", *A New World Record*, ELO   
http://home.earthlink.net/~yatescr

dbell wrote:
> All,
> 
> This is for real.
> 
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
> 
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
> 
> What can you say about the device performance?

It means that it's 1.8 dB worse than the limit set by quantizing noise. 
For full-scale input, there is 96 dB SNR and no headroom. Reduce the 
signal level by 20 dB and you get 76 dB SNR and 20 dB headroom. Assuming 
perfect linearity and not too close to zero signal, the sum of headroom 
and SNR will be 96 at any level.

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

All,

A number of people who responded made valid points (you can reread the
posts), including what I was going to discuss. The spectrum was
computed with an FFT, type of window I don't remember.  It DID show a
MINIMUM spur-free range for the frequency present because no spurs were
visible. The apparent intention (my conclusion) of the advertizer was
to show near perfect 16-bit performance, which it DID NOT show.

One spectral signal point above one spectral noise point by 96 dB is 96
dB SNR, but there were MANY spectral noise points.  If the frequency
was right on a bin and had no windowing and the FFT size was
approximately 2048 (with no FFT mag averaging), there was a single
spectral signal point and approximately 1000 spectral noise points (all
about 96 dB below the signal), so the SNR would be about 96-30=66 dB
which would be REALLY POOR PERFORMANCE for a 16-bit A/D.

Dirk

dbell wrote:
> All,
>
> This is for real.
>
> A 16-bit A/D manufacturer showed a spectral plot to indicate the
> performance of their device.
>
> The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> vs frequency (Hz). There was a fairly flat noise floor (grass if you
> prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> frequency expected.  The peak was 96 dB above the noise floor.
>
> What can you say about the device performance?
> 
> Dirk
> 
> Dirk Bell
> DSP Consultant

hi dbell,
          can you please explain the last paragraph, still more
indepth. i really didnt understand.
please

regards
particlereddy


>
> A number of people who responded made valid points (you can reread the
> posts), including what I was going to discuss. The spectrum was
> computed with an FFT, type of window I don't remember.  It DID show a
> MINIMUM spur-free range for the frequency present because no spurs were
> visible. The apparent intention (my conclusion) of the advertizer was
> to show near perfect 16-bit performance, which it DID NOT show.
>
> One spectral signal point above one spectral noise point by 96 dB is 96
> dB SNR, but there were MANY spectral noise points.  If the frequency
> was right on a bin and had no windowing and the FFT size was
> approximately 2048 (with no FFT mag averaging), there was a single
> spectral signal point and approximately 1000 spectral noise points (all
> about 96 dB below the signal), so the SNR would be about 96-30=66 dB
> which would be REALLY POOR PERFORMANCE for a 16-bit A/D.
>
> Dirk
>
> dbell wrote:
> > All,
> >
> > This is for real.
> >
> > A 16-bit A/D manufacturer showed a spectral plot to indicate the
> > performance of their device.
> >
> > The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> > vs frequency (Hz). There was a fairly flat noise floor (grass if you
> > prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> > frequency expected.  The peak was 96 dB above the noise floor.
> >
> > What can you say about the device performance?
> > 
> > Dirk
> > 
> > Dirk Bell
> > DSP Consultant

Hi dbell,

first of all, thank you for the interesting question - it scratches one
topic I tried to understand completely some weeks ago, but I had to
leave it for other work... Looks like you and the others (hello, by the
way) could be the perfect discussion partners for the questions raised.

It is about specifying spectra as power spectra or power spectral
density results. Assume a system which emits a number of tones, i.e.
single frequencies within the spectrum. There is also broadband-like
noise in the system, i.e. some pink-shaped noise, and broadband noise
with white characteristics, e.g. from fans or background noise.
Assume further I could measure the sources in sequential order, i.e.
just noise A, then A+B, then A+B+C and so on. I would like to compare
the sources, and specify e.g. the SNR of the different sources with
respect to the lowest (white) noise.

I could show it in a frequency domain plot as power spectrum or
magnitude spectrum, but depending on the number of FFT-coefficients,
the broadband noise will be raised or lowered, while the single-tone
signals will keep their values. This is why for broadband noise one
will typically use the power spectral density, which gives a definite
number, independend of the FFT used. Nevertheless, defining power
spectral density (V^2/Hz vs. Hz) for single tones not only feels
somehow weird, it also for these signals gives results which will be
lower or higher depending on the number of FFT coefficients.

My question is: how do I state a comparison value like SNR for such a
case (mix of different broadband noises and single tones), and how can
I show it graphically in a way that makes sense: Power spectrum or
power spectral density?

Any thoughts are most welcome. I didn't found a solution which solves
all the questions raised.

stereo



dbell wrote:
> All,
>
> A number of people who responded made valid points (you can reread the
> posts), including what I was going to discuss. The spectrum was
> computed with an FFT, type of window I don't remember.  It DID show a
> MINIMUM spur-free range for the frequency present because no spurs were
> visible. The apparent intention (my conclusion) of the advertizer was
> to show near perfect 16-bit performance, which it DID NOT show.
>
> One spectral signal point above one spectral noise point by 96 dB is 96
> dB SNR, but there were MANY spectral noise points.  If the frequency
> was right on a bin and had no windowing and the FFT size was
> approximately 2048 (with no FFT mag averaging), there was a single
> spectral signal point and approximately 1000 spectral noise points (all
> about 96 dB below the signal), so the SNR would be about 96-30=66 dB
> which would be REALLY POOR PERFORMANCE for a 16-bit A/D.
>
> Dirk
>
> dbell wrote:
> > All,
> >
> > This is for real.
> >
> > A 16-bit A/D manufacturer showed a spectral plot to indicate the
> > performance of their device.
> >
> > The spectrum was from digitizing a single tone, plotted Magnitude(dB)
> > vs frequency (Hz). There was a fairly flat noise floor (grass if you
> > prefer) from 0 to Fs/2.  There was a single peak in the spectrum at the
> > frequency expected.  The peak was 96 dB above the noise floor.
> >
> > What can you say about the device performance?
> > 
> > Dirk
> > 
> > Dirk Bell
> > DSP Consultant