I agree. It’s my fault I’m not describing this clearly, but I was referring to switching between digital and analog signals. One can do math using analog circuitry and the interpret the result as digital. The basic problem that I’ve been hinting at is that what we call “digital” has analog components, and what we call “analog” in computing usually turns into digital eventually. (Maybe not analog audio processing, but analog computing does.) It’s often a mix, and there’s a big gray area.
> there are endless debates about whether class-D amplifiers are “digital” or not.
Right! I believe there are debates about whether modern “analog” audio delay pedals are analog or digital too, right? They are now using more cheap digital components to simulate analog delay lines, which is really if you think about it switching between analog and digital multiple times before you ‘hear’ it. Does it even make sense to summarize it as either analog or digital?
> “Noise” is ill defined by itself. It means an unwanted fluctuation of a signal. [..] The vagaries of floating point arithmetic [..] are interesting but basically sideshows.
I guess it’s getting lost, but I’m making an analogy between noise and floating point, not claiming that FP rounding error is the same thing as noise. I’m perfectly familiar with noise, and I’m just saying that, in a way, floating point rounding isn’t entirely dissimilar from the net effects of noise: they both are accurate to within a certain tolerance, and they both grow larger errors with chained operations. Fp32 is way more accurate than a typical analog signal, but that’s because it has a lot of bits. Fp8 on the other hand might really have closer to the accuracy that a high quality analog signal has. Digital fp8 might be repeatable where analog isn’t, but that doesn’t necessarily make it more accurate, nor is it something we characterize as “exact” like @corethree was claiming.
My only point in this thread was to say analog computations do have some uses, which is why they’re still being used and studied. @corethree said there’s no practical use because analog can be simulated with digital. I’m sure as an audiophile, you’re aware that claim isn’t true. It’s true that analog can be simulated with digital but that’s a slow simulation and has no bearing on whether analog components are useful.
One of the analog computing applications I had in mind when I replied was the new Photonic chips. https://www.nature.com/articles/s41586-023-06558-8
Another one is quantum computers, they’re quite analog, and definitely different from ‘digital’.
A third is fast low-precision AI training chips. There are a whole bunch of startups trying to make these come to market right now, all of them full of smart people who’ve studied the problem and believe that analog computation for AI training solves a real problem.