FWIW, it seems I've found a phenomenon with the AK4490 (of my RME Adi-2 Pro FS) that might be some form of noise-floor modulation. Between around -30dBFS and -50dBFS I see an increase of harmonics and mirrored harmonics in the upper frequency range. If one were to factor in these in a wideband THD (sans noise) measurement we would have a mid-level THD rise (not unlike the ESS chips) even when the low components stayed completely the same, because at higher levels the HF components get lower and lower in level and get spread out in frequency range, and at lower levels they get shifted higher and higher in frequency, hence making less of a contribution. Actually, only a hint of H2 was visible, and always lower in level than this new dirt.
See attached plot, a loopback measurement (348kHz SR, 512k FFT, 32 averages). Shown is a 300Hz signal at levels from -34dBFS to -42dBFS. Traces have been separated by 10dB for visibility, center trace has no offset. ADC runs at 20dB gain to eliminate its influence (DAC level +24dBu, ADC level +4dBu). ADC is not the culprit, double checked with analog generator (Tek SG505). Things do *not* depend on sample rate. With other rates than 384kHz used here only the right side of the plots is truncated, otherwise no change. Frequency also doesn't matter much (and don't need to be bin-centered. If it's not a bin-center, FFT needs window which raises processing noise floor). With higher freq the spacing gets wider and the onset of the mirrored components is more readily visible (see below). Wider spacing means less total energy, therefore the effect is worse at lower frequencies.
The really interesting thing is that the error spectra get heavily shifted around in frequency, especially the start of the error components, with rather moderate level changes of 2dB.
Also, we can see at -42dBFS the needles are all equally spaced at 300Hz intervals. At the higher levels, as second train of needles appears in the right-hand area of the plots which are 300Hz-spaced, but not on 300Hz multiples -- looks like mirrored at the sampling frequency (clearly visible when zoomed in and/or using higher test freq).
Those tones I saw can actually go way below 20kHz with some combinations of frequency and level, so it doesn't seem AKM did some form of shaping to have those tones strictly above the audio band, or at least it is not fully effective.
Not shown, I made further observations: If the test signal is not a pure sine, components start to reduce quickly. Even when a second tone (of arbitrary frequency, but preferably lower and unrelated) is 60dB down to the main tone, the hash disappears. DC level send to the DAC matters as well (which puts the modulator into a different operating point), as soon as the DC is higher than ~ -30dBFS, the needles go away).
Of course, all the dirt is way below RMS noise floor, also we don't have 100% pure sine content in any music, so it's possibly completely inaudible... but actually we don't know...
View attachment 30501
(also posted on DIYaudio.com a few days back)