[EchoAce]

It may be my lack of knowledge about optics, but from an ML perspective this seems rather mundane if not useless. Model training involves high levels of parallelism on a large scale for difficult tasks, something I can’t see these optical chips doing. Does anyone have any further information that might enlighten me to otherwise?

[stochastic_monk]

By performing these transformations optically, they primarily get data parallelism (like [GTV]PUs). I expect this to happen. NVIDIA’s ACDC paper provides an FFT-accelerated neural network layer (similar to deep-fried convnets), with an offhand remark that the transformations could be performed optically. I wonder what kind of information bandwidth they can get, though.

[dekhn]

Physicists were using optical lenses to do approximate FFTs over a hundred years ago.

[drumttocs8]

Can you use something like wavelength division multiplexing to get different data streams and achieve parallelism there?

[d--b]

I can't read the article as it's behind a paywall, but if you can make a chip that's 100% optical, then it means that when you beam your input data at the input end of the chip, you _instantly_ get the output at the end. No need for cycles for multiplying, adding and so on. Plus it wouldn't heat up like silicon does.

[p1esk]

you _instantly_ get the output

That's not how physics works, unfortunately.