4 ms·

AVX-512 software is appearing. For example, for neural net inference: https://NN-512.com https://NN-512.com AVX-512 is a pleasure from the programmer's perspec

AVX-512 software is appearing. For example, for neural net inference: https://NN-512.com https://NN-512.com AVX-512 is a pleasure from the programmer's perspective, much nicer than writing software for GPUs or for earlier SIMD instruction sets like AVX2/SSE Versus a GPU, the AVX-512 machine model is just much simpler, and there are no drivers or proprietary compilers Versus earlier SIMD instruction sets, AVX-512 provides all the instructions you want (permutations, floating point conversions, etc.) and they're all masked so edge cases are easy. AVX-512 Foundation is a really nice instruction set and it has a bright future

dogma1138 6y ago
Isn’t the issue with AVX-512 is that different CPUs support different instructions? https://en.m.wikipedia.org/wiki/AVX-512 https://en.m.wikipedia.org/wiki/AVX-512 It might eventually get normalized but atm it’s a bit of a mess.
- wtallis 6y ago
  It's a bit less messy if you ignore the GPU-like Xeon Phi (Knights Landing and Knights Mill) and the failed Cannonlake that Intel definitely would prefer everyone pretend never happened. But that still leaves you with a progression of four levels of expanding support, and the Cooper Lake mess of having BFLOAT16 but not the stuff added in Ice Lake and Tiger Lake. If you don't cut down the support table to eliminate irrelevant niche products, it's most of the way to being a Gray code.
  - dogma1138 6y ago
    I think Icelake and Tiger lake not having BF16 is going to be a big issue especially for machine learning applications, I can’t find what Rocket Lake supports yet but I’m guessing it’s going to be similar to Tiger Lake. There is also the big unknown of what kind of support AMD will bring to the table. I think Intel is banking on people using OneAPI more than writing AVX-512 code using intrinsics or compiler magic. At that point compatibility might be not as much of an issue if you ignore potentially drastic differences in performance.
- 37ef_ced3 6y ago
  If you limit yourself to the Foundation instruction set (AVX-512F) you get a lot of benefit and maximum compatibility. That's what https://NN-512.com https://NN-512.com does

dogma1138 6y ago

Isn’t the issue with AVX-512 is that different CPUs support different instructions? https://en.m.wikipedia.org/wiki/AVX-512 https://en.m.wikipedia.org/wiki/AVX-512 It might eventually get normalized but atm it’s a bit of a mess.

wtallis 6y ago
It's a bit less messy if you ignore the GPU-like Xeon Phi (Knights Landing and Knights Mill) and the failed Cannonlake that Intel definitely would prefer everyone pretend never happened. But that still leaves you with a progression of four levels of expanding support, and the Cooper Lake mess of having BFLOAT16 but not the stuff added in Ice Lake and Tiger Lake. If you don't cut down the support table to eliminate irrelevant niche products, it's most of the way to being a Gray code.
- dogma1138 6y ago
  I think Icelake and Tiger lake not having BF16 is going to be a big issue especially for machine learning applications, I can’t find what Rocket Lake supports yet but I’m guessing it’s going to be similar to Tiger Lake. There is also the big unknown of what kind of support AMD will bring to the table. I think Intel is banking on people using OneAPI more than writing AVX-512 code using intrinsics or compiler magic. At that point compatibility might be not as much of an issue if you ignore potentially drastic differences in performance.
37ef_ced3 6y ago
If you limit yourself to the Foundation instruction set (AVX-512F) you get a lot of benefit and maximum compatibility. That's what https://NN-512.com https://NN-512.com does