6 ms·
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 paralle
by EchoAce 8y ago
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?
- d--b 8y agoI 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 8y agoyou _instantly_ get the output That's not how physics works, unfortunately.
- stochastic_monk 8y agoBy 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 8y agoPhysicists were using optical lenses to do approximate FFTs over a hundred years ago.
- drumttocs8 8y agoCan you use something like wavelength division multiplexing to get different data streams and achieve parallelism there?