There is a lot here.
A NOS DAC has no special capability wrt exact voltages, it suffers from exactly the same issues as a DAC running at any other rate.
Right, what I meant is that (based on my certainly incomplete knowledge) with a NOS DAC you know based on the digital input signal and the DAC's voltage range what voltage states it would ideally transition between, and hold for a considerable amount of time. In essence, if you briefly measure the voltage levels of an ideal NOS DAC at the right interval (sample its output signal at the input bit depth and sample rate), you should get the same data as you put in. Maybe that assumption is already wrong, but if not at least there's some objectivity there, and resistor ladder based DACs may already struggle with this depending on the bit depth because you need resistors with higher and higher precision, or so I hear. I'm assuming that's what Amir's linearity tests are ultimately about.
The next easiest thing in theory would be an oversampling DAC with a closed form filter that like Schiit claims (for its multibit DACs) preserves the original samples (and only adds new ones inbetween). If you sample the output signal at exactly the right times (ignoring the added samples), you should ideally see exactly what an ideal NOS DAC would do, namely the input samples - again, based on my limited knowledge. It's trickier here because the higher the internal sample rate, the less time there is to reach and hold a certain voltage level.
With any other DAC you would need to know the exact filter that it uses to even know what voltage states it is trying to achieve in order to judge how accurately it is doing so - if you wanted to simply judge it based on individual samples rather than something more accumulative like an FFT.
That's what I mean by lack of objectivity - there is no predefined ideal voltage level at certain times without knowing the filter. If you have a simple, single short impulse and see much longer ringing, the DAC is clearly taking some liberty in how it interprets the input, injecting samples even before there was any signal at all. You cannot judge how accurately it is doing so without knowing what it is trying to produce.
The core issue is that no real DAC creates an infinitesimally short output pulse. However the ideal ADC samples at an infinitesimally short time.
This is akin to low persistence displays, right? In virtual reality displays for instance, you want a very short, bright pulse instead of a longer, less bright image because the latter results in a smearing sensation as you move your head, since you see exactly the same thing despite a change in perspective. A low persistence display addresses this by not presenting any stimulus most of the time (and increasing the brightness when it does show something to compensate).
And in sound reproduction, if there is no continued pressure change for a certain time, and then a sudden one, followed by another lack of a change (to the extent physics allows), it's like seeing the same image while your head moves. Time passes, but the stimulus doesn't change. But for sound there is no equivalent to the darkness in between bright frames, correct? Essentially, I cannot avoid an acoustic stimulus between samples.
An ideal DAC with output filter creates sinc function pulses, and these overlapping sincs sum to the precise bandwidth limited input the ideal ADC sampled. A sinc function is of infinite length. So in theory, you need to sum every one of the sinc functions at every sample point in order to assemble the signal, which for practical purposes means you are going to need a long latency. Even 600ms isn't enough. Everything is an approximation.
I definitely only understand a tiny bit of that. Maybe some day. This dude wants to make some videos about how different DACs work, maybe he is able to spoon feed me what would take me much longer to gather from a text book:
https://youtube.com/c/GoldenSound.
However the question then becomes - does it matter? How quickly do these effects vanish below the noise and are impossible to actually find? A lot of this naval gazing is done without considering the impact of noise, something that is an intrinsic part of the real universe rather than the one where such obsessing exists. A sinc function dies away very quickly. This is important. 600ms is actually insanely over the top. There is no useful information out there and the whole idea is just bragging rights. The limits to the amount of information actually present are clear, and in reality, it just doesn't matter.
It's clear that we can reconstruct meaningful information from sound (say, speech) with very noisy, distorted reproductions. And that we can recognize what kind of instrument something is supposed to be even with pretty bad recordings and/or reproductions, despite knowing that it's just a played back recording. But it's really hard to judge how far one has to take it to completely fool human hearing into a sense of presence that is indistinguishable from reality. My impression is that, say, Rob Watts and Mike Moffat are after that, for instance, not bragging rights.
You need a reconstruction filter, that is an absolutely iron clad reality. What is also iron clad is that real life filters are never ideal. There is always a compromise. And that compromise is fluid between the digital and analog domains, giving the designer the freedom to do implement in the most effective domain.
Right. I accept that, though with some disappointment, because it means digital audio is unable to fulfill its promise of perfection, even before we get into the messy analog realm, which is inherently imperfect.
It's a little like how I feel about Bluetooth, a new and improved version every so often, but what I really want is lossless transmission and be done with it, but then they wouldn't have to sell us a newer, fancier model next year. At least with DACs, it's simply impossible rather than a business decision.
Of course there is also no perfect ADC, but then again not all of digital audio is recorded.
The NOS adherents don't really get this, and they somehow believe that everything possible should be done in the analog domain when the modern reality is that we have such huge capabilities available in the digital domain that is makes no sense to do things in the analog domain if you can avoid it. So NOS DACs are intrinsically compromised by the limitations possible implementing a real world analog reconstruction filters that work in the limited bandwidth they have to spare. Or worse, they somehow think that huge amounts of aliasing into the audio band is yielding a better result and they compromise the reconstruction filter.
Probably true for many of them, but I appreciate those that like the thought of a Holo Audio NOS DAC with 1.536 MHz PCM support via USB - so that you can go bananas with filters in software, for instance via HQPlayer, instead of layering two filters (a software one and whatever the DAC forces on you).
