3 ms·
Actually, a GPU is not the best kind of hardware for AI (or rather, neural network simulation). It's very well-developed, and more useful than a CPU, but one co
by einpoklum 3y ago
Actually, a GPU is not the best kind of hardware for AI (or rather, neural network simulation). It's very well-developed, and more useful than a CPU, but one could design and construct more NN-sim-oriented hardware. NVIDIA have been forced to do this somewhat, cannibalizing some of their more general-purpose compute capabilities in favor of matrix-multiply-add functionality (i.e. "tensor cores"). That's not exactly the "GPU way" of doing things. And one could go every further, perhaps all the way up to some analog computations instead of digital low-precision ones.
- cmrdporcupine 3y agoYeah I'm confused why e.g. Google's TPUs, etc. didn't catch favour in the market? I'm not an ML developer, so I'm curious. Just economies of scale?
- kimixa 3y agoThere's a natural compromise between "Efficiency" and "Flexibility" - training AIs is currently a fast enough moving field that by the time the really efficient targeted hardware could be released, the state of the art has moved to make them less suited to the latest generation. Outside of training, using the trained network for inference seem to be changing less, so maybe a decent target for bespoke targeted hardware. And indeed it seems that custom hardware blocks are very much "catching favor" in that market - everyone seems to be adding accelerators to their SoCs. GPUs seem to be at a happy medium, where they're often flexible enough to run techniques while still having better efficiency than a CPU.
- xxpor 3y agoWell the first issue is random researchers can't just buy a TPU and throw it in their workstation. (No, corals don't count)