Re:Linux founder tells Intel to stop inventing magic instruc...
by cogman10 on Monday July 13, 2020 @03:32PM
(#23824664)
I think you are thinking AVX takes up more diespace than it does. Only a tiny fraction of all CPU diespace is dedicated to the actual CPU. The rest goes primarily to cache and partially to the various controllers on the CPU (memory controller, PCI-Express controller, etc).
There's a reason AMD is and Intel are able to cram full GPUs onto the same die as the CPU.
Re:Linux founder tells Intel to stop inventing magic instruc...
by Dylan16807 on Monday July 13, 2020 @03:59PM
(#23824949)
https://imgur.com/r/intel/4SG36
Yeah, here's a good image of how much space it costs to add a 'second' AVX-512 unit to certain SKUs. Cutting AVX-512 support entirely would save a bit more space, and the total would be enough to make a difference, but it's not an extreme amount.
https://software.intel.com/sites/default/files/managed/5f/99... https://twitter.com/GPUsAreMagic/status/1256866465577394181 these too (speculative)
The only AVX512 instructions that actually seem essential imo are the VNNI extensions.
There are a lot of use cases for a neural network that is trained on the GPU but where inference is done on a bunch of CPUs because it's cheaper than buying/renting inference hardware.