Hacker News Reader

Comment by adamnemecek | Hacker News Reader

They are much faster, also can solve optimization problems very naturally.

Analog computing used to be a thing before ww2. It was dope for numerical computation but came out of fashion due to rift. Photonic analog quantum computing is the future.

View on HN

Analog computing has problems of accumulating error that can’t be resolved.

The more sequential operations you have to perform, the more the error accrues because unlike any discrete system you have no ability to filter noise.

That’s why analog computers “fell out of fashion”. That and the relatively restricted programmability, higher power usage, etc.

There are some problems where analog can be faster, but the moment you have consecutive operations you need to accept imprecise answers.

Yes, that’s rift. Photonics don’t have that problem.

When I search ddg with the following query:

    rift "analog" "computing" -oculus

The only relevant result is your comment from this thread.

Accumulated error creates an gap, rift between the true answer and the computed answer.

Something I've never been able to figure out is whether our modern, deep sub-um and even nm-scale process nodes, are directly usable for anything analog. I suspect that they aren't - that the noise levels are now at a point where even the simplest "gate"-like structures are only useful if used in combination with that "attraction/iteration towards a fix point" mechanism that essentially defines digital electronics (from an "analog" POV), with its inherent resilience to noise. Unfortunately, I even doubt that the question is directly answerable without the sorts of proprietary data that only a multi-billion chip fab would have in the first place.

They are, certain components even in the smallest process have to be analog. One good process for analog functionality is 22nm SOI, 65nm TSMC also works very well.

What happens to this rift if it's operating on an equation with a derivative such as the gradient of a neural network?

I'm a huge fan of deterministic computation, but we seem to be doing okay without it in deep learning with respect to floating point round off error. Or despite the fact that the computation could be deterministic, the algorithms are written in a way that they are not due to asynchronous accumulation and computation.

Would we just get a different minimum each time we run or would something more pernicious occur?

Reply on news.ycombinator.com