JohnRoberts said:
Be aware that the shape of the noise floor around the signal is an artifact of the FFT calculation and windowing(?) decisions.
Yes, that's why I made sure the settings would be the same for all the analyses, and indeed, it qualifies only for comparative assessment. Since I used internally generated files that start and end with a zero-crossing, I could have dispensed with windowing, but I thought it was not necessary, considering the FFT was stable enough after 3-4 seconds.
I repeat that old school (analog) bench THD+N and S/N measurements can be instructive.
Remember the issue was raised about the intrinsic capability of higher bit depth to produce better performance, so bypassing any converters adresses the issue at the heart.
Indeed, what's important is what comes out of the orifices, and analog measurement is a proven evaluation method; however, we're dealing here with differences that most audio analysers would not register. AP555, the unquestioned champion, has a residual THD of -117dB and a self-noise about the same. With it, one can certainly assess the difference between 16 and 24 bit, but not between 24 and 32. And anyway, whatever analog circuitry in the converters would tend to mask any difference. And not everybody has access to a 555, or even a 525.
I believe no one in his right mind can deny that 24bit gives superior performance to 16 bit, provided the implementation is correct. There is enough proof that well-implemented converters do exist, delivering performance close to the chip mfgrs specs.