Using Cross Corelation to lower influence of ADC for DAC measurements

[KSTR]

So what is the bug, when the level is correct? Are we arguing about -28.86 vs -28.90?

 

[KSTR]

@FrenchFan ,
You measured a real DAC with a real ADC, and both have gain and clock instabilities/drifts and noise, therefore a 0.04dB level difference is absolutely well within tolerances of general repeatability.
If you're searching for processing "bugs", please use reproducible (synthesized) signals.

 

[Sokel]

What I have seen during testing these days is that a lot of CCs make nice charts but don't change results much after the first 1k or so CCs:

Here's with 1k and 5k CCs.


1k


5k

(It's amazing that I'm within 1db from AP and the same distortion structure,etc with my ancient interface! )

 

[FrenchFan]

@FrenchFan ,
You measured a real DAC with a real ADC, and both have gain and clock instabilities/drifts and noise, therefore a 0.04dB level difference is absolutely well within tolerances of general repeatability.
If you're searching for processing "bugs", please use reproducible (synthesized) signals.

KSTR

Yes I agree regarding the signals for
reproducibility.

I just wanted to show
only for a given condition with "my" DAC
and "my" ADC we couldn't be proud of
RO+ result and this completely
random in my eyes.

I tried this morning with a signal containing
3 lines at -140dBFS and RO+ worked perfectly.

That's all I wanted to show.

 

[KSTR]

I just wanted to show
only for a given condition with "my" DAC
and "my" ADC we couldn't be proud of
RO+ result and this completely
random in my eyes.

Once again, what is "random" about a truly irrelevant 0.04dB difference?

 

[KSTR]

@Sokel,
When too much bins disappear I wouldn't trust the results, notable the noise estimation will be too optimistic.

 

[Sokel]

@Sokel,
When too much bins disappear I wouldn't trust the results, notable the noise estimation will be too optimistic.

I know the results are not to be trusted even if they are repeatable all the time.
Noise is some 10dB down,that's too good to be true,that's what I get normally:



Harmonics are close but noise

 

[Sokel]

@Sokel,
When too much bins disappear I wouldn't trust the results, notable the noise estimation will be too optimistic.

Also tried with longer FFT to see and I get pretty much the same:





Low bins disappear at the first 100 CCs or so,the rest up high takes longer.

 

[KSTR]

Low bins disappear at the first 100 CCs or so,the rest up high takes longer.

Bins should never disappear during the ongoing correlation process, quite the contrary, they should reappear once we're closing in on the DUT noise floor.
When bins disappear that means you have some correlated but out-of-phase content.
With test signals I could verify that as little as 3% out-of-phase content is critical. I've used de-correlated noise in both channels, then applied 3% inverting cross feed and bingo, bins start to disappear with R0+ CC averaging

 

[FrenchFan]

Once again, what is "random" about a truly irrelevant 0.04dB difference?

0.04dB it is not a problem , is BW, it stop at 10KHZ , it is in HF the PB

 

[KSTR]

Harmonics are close

But they should be the same, as H2/H3 are far above the FFT noise floor.
Something is fishy...
I've verified that REW is really spot on, when looking at the peaks in the waveform... of course you have to de-embed the FFT noise floor manually when you are just a few dB's above.

 

[KSTR]

0.04dB it is not a problem , is BW, it stop at 10KHZ , it is in HF the PB

Explained in post #309, you seem to have some out-of-phase content in the top end.
And as noted in previous post of mine, one should always double-check with all methods available to sort out things, don't blindly trust what you get.

 

[KSTR]

of course you have to de-embed the FFT noise floor manually when you are just a few dB's above

Details on this here: https://www.edn.com/measuring-small-signals-accurately-a-practical-guide/
It also deals with CC ;-)

 

[Sokel]

Bins should never disappear during the ongoing correlation process, quite the contrary, they should reappear once we're closing in on the DUT noise floor.
When bins disappear that means you have some correlated but out-of-phase content.
With test signals I could verify that as little as 3% out-of-phase content is critical. I've used de-correlated noise in both channels, then applied 3% inverting cross feed and bingo, bins start to disappear with R0+ CC averaging

Oh,you talk about 2-channel.I should have state earlier that this is one channel into the 2 ADCs,with a splitter.
Or you mean that there are phase difference between the ADCs?
Here's E-MUs diagram:



I use line-in inputs of course,R ch of Tone Board to L R ch of E-MU,unbalanced.
At REW,harmonics are identical depending the window but within 1dB,as with Multitone.113-114dB for H2 and H3),but it's handy to use Multitone for normal ones as for now until CC averaging is implemented there too.

 

[Sokel]

But they should be the same, as H2/H3 are far above the FFT noise floor.
Something is fishy...
I've verified that REW is really spot on, when looking at the peaks in the waveform... of course you have to de-embed the FFT noise floor manually when you are just a few dB's above.

Did a silly test to isolate the problem.
I removed all connections,used ASIO instead of Exclusive and let E-MU run idle without signal averaging it's ADCs with CC.
That's what I got:




So,the problem is there,if I understand well what you said,ADCs have phase difference (?) that causes the bins to disappear?
(no wonder,we're talking 20yo interface!)

 

[KSTR]

Oh,you talk about 2-channel.I should have state earlier that this is one channel into the 2 ADCs,with a splitter.
Or you mean that there are phase difference between the ADCs?

The latter, one-channel source signal to two ADCs via splitter.

So,the problem is there,if I understand well what you said,ADCs have phase difference (?) that causes the bins to disappear?
(no wonder,we're talking 20yo interface!)

Test run with three signals, 700 CC R0+ Averages.
- fully decorrelated channels
- -3% cross-feed, so creating some small out-of-phase content
- +3% cross-feed for some in-phase content

As soon as we have out-of-phase signal, more bins start to disappear. OTOH, even just a bit of fully correlated signal makes less zero bins. The more averages, the stronger the effect.
And for fully de-correlated signals, about 50% of bins go to zero (better visible with linear frequency scale).

IMHO this strongly indicates when many bins have gone to zero there must have been some out-of-phase content in the noise, at least I have no other explanation for it (but there might be one).
One way to check would be to use "inverted correlation reference" option. If that gives you a plot with fewer zero bins then you know there is out-of-phase stuff (which is now fully correlated, in-phase)

 

[Sokel]

The latter, one-channel source signal to two ADCs via splitter.

Test run with three signals, 700 CC R0+ Averages.
- fully decorrelated channels
- -3% cross-feed, so creating some small out-of-phase content
- +3% cross-feed for some in-phase content
View attachment 387555
As soon as we have out-of-phase signal, more bins start to disappear. OTOH, even just a bit of fully correlated signal makes less zero bins. The more averages, the stronger the effect.
And for fully de-correlated signals, about 50% of bins go to zero (better visible with linear frequency scale).

IMHO this strongly indicates when many bins have gone to zero there must have been some out-of-phase content in the noise, at least I have no other explanation for it (but there might be one).
One way to check would be to use "inverted correlation reference" option. If that gives you a plot with fewer zero bins then you know there is out-of-phase stuff (which is now fully correlated, in-phase)

Took me a while but I think I got it fixed:



Thanks for that!

 

[nanook]

Took me a while but I think I got it fixed:

View attachment 387695

Thanks for that!

Maybe I'm on the hose... What did you do differently to get this result?

 

[Sokel]

Maybe I'm on the hose... What did you do differently to get this result?

This is a repeat of the #315 post one,with no signal and no connections,just an CC averaging of the noise essentially.
I didn't do anything different considering the measurement,what I did though was a little cleaning,re-soldering,changed some caps around the inputs,etc.
It was the right channel ADC who that had the problem and KSTR got me suspected of it.

The stupid thing of my part is that I didn't thought to measure the R Ch of it's own in depth as it appeared to be ok,but...
Another thing that can make it suffer is its low input impedance,so long cables,etc is a big no.

 

[nanook]

Test run with three signals, 700 CC R0+ Averages.
- fully decorrelated channels
- -3% cross-feed, so creating some small out-of-phase content
- +3% cross-feed for some in-phase content

It would be interesting to see the "cross-correlation magnitude" in comparison.
My expectation is that it ends up at fairly the same level as the "+3% crossfeed" but with almost no bins going zero.