[solved] RME Adi-2 Pro FS (AK4490 DAC): positive sample values are always offset by -1 !

[KSTR]

A few days ago RME has released a beta version of the upcoming firmware revision which fixes the problem, and I could confirm that by repeating the initial measurement in post #1.

Now the staircase waveform is reproduced faithfully, with full symmetry:

As an additional check of digital signal integrity I also recorded the RME's response with the local DAC level setting at -15dBr, which means the signal must be rendered with dithering to obtain mean values that are not LSB multiples anymore. The block averaging also reduces the dither noise and this is what we get:

Even though we're talking sub-LSB levels here (-15dBr digital DAC gain means one input bit is ~18% of an output LSB), those steps are still visible and have the correct relationships. The first "-1" value (cursor) looks a bit weak and the DC offset for "0" looks to be slightly negative, but now we are approaching ADC limits and this little error, in case it really came from the DAC and pre-processing, would be truly academic anyway.

From this results I would conclude that the preprocessing in the RME Adi2- Pro is 100% flawless now.

The internal resolution, needed for proper digital level reduction and re-dithering for the actual DAC chip input also showed to be fully adequate (and probably has always been), though I'm not clear about the significance of this measurement.

Big thanks to RME and Mathias for taking action to remove this little processing glitch.

 

[KSTR]

I did a comparison of old vs. new firmware using more complex multitone signals and the good news is that even with the offset error I could not see any difference in the spectra when measured normal mode, non-averaged. The additional error terms never make above the analog noise floor and this was already seen in the single sine measurements. This again makes clear that this is a *very* low level error which IMHO is most certainly inaudible.

1) 6-tone signal with each component at -40dBFS:

red: non-averaged response, spectrum is clean down to the noise floor.
yellow: block-averaged (40dB noise reduction) *without* offset error. Some IMD components clearly stick out now, at the expected frequency points.
blue: block-averaged (40dB noise reduction) *with* offset error. It's a bit hard to see but at no point the magnitude are any lower whereas there is much more "grass" growing on the floor. This is correlated and wideband so it's correlated noise, basically.
Note: the ripple at LF comes from the block size being only 4409 samples (~100ms, 10Hz) which has been looped 16 times in order to apply a 64k FFT. Only multiples of 10Hz can generate content, therefore.


2) same, but now at -80dBFS component level, and linear scale:

yellow: non-averaged response, spectrum is clean down to the noise floor.
red: block-averaged (40dB noise reduction) *without* offset error. Spectrum still is as clean as it gets (and that's very nice).
blue: block-averaged (40dB noise reduction) *with* offset error. The wideband overall effect of some 10dB additional correlated noise is evident. The analog noise floor still is some 20dB higher and will surely dominate, the added hash unlikely to be a problem in any other case than this specific sub- noisefloor type of measurement.

It was an unexpected but interesting exercise for me because I never dealt with this type of very subtle digital error before and it turned out that sample-synced block-averaging is a very mighty tool in the measurements box. It has restrictions, of course...

 

[pos]

Wow, nice job from start to finish