The instruction set was actually announced three years earlier, in 2013, presented by Intel as “a significant leap forward towards 512-bit SIMD support. Programs can contain eight double-precision or sixteen single-precision floating point numbers, or eight 64-bit integers or sixteen 32-bit integers within 512-bit vectors. This allows processing of twice the number of data items that Intel AVX / Intel AVX2 can process with a single instruction and four times that of SSE. Intel AVX-512 instructions are important because they offer more performance for the most demanding computing operations. ”
Torvalds mentioned it on a mailing list that commented on the apparent lack of support for AVX-512 by future Alder Lake and Sapphire Rapids architectures. Torvalds justifies his invective in an articulated way: “I hope AVX-512 dies of a painful death and Intel starts solving real problems instead of trying to create magic instructions and then create benchmarks in which they look good. I hope Intel goes back to basics, focusing more on the regular code that is not related to HPC or to some useless special case“.
“I have said this before but I will say it again: in the heyday of x86, when Intel killed all the competition, everyone did better than Intel with FP loads. Intel’s performance with those loads was worthwhile – relatively speaking “And nobody cared. That’s because nobody cares about it outside the benchmarks. The same is largely true of the AVX-512, now and in the future. Yes, you can find aspects that interest you, but they are not things that do not sell.” computers in the big picture “.
“AVX-512 has real disadvantages. I’d rather like to see that transistor budget used on other things that are much more relevant, whether it’s still floating point calculations – in the GPU instead of AVX-512 – or more cores with good single-thread but no junk performance like AVX-512, like AMD did. ”
“I want to reach the power limits with the regular integer code, not with some energetic vampire like AVX-512 that cuts the maximum frequency (because people end up using it for memcpy!) And reduces the cores (because those useless garbage units take up space ) “.
“Yes, yes, I’m biased. I absolutely hate the benchmarks on floating point calculations and I understand that others are deeply interested in them. I just think the AVX-512 is exactly the wrong thing to do. […] It is an excellent example of something Intel has done wrong, partially increasing market fragmentation. No more junk for special cases and make all the common things that everyone cares about work. So make an FPU that’s just pretty good and people will be happy. AVX2 is much more than enough. ”
Intel has not yet responded to Linus’ dad’s words and will hardly do so. Torvalds recently switched to its AMD hardware main system after Intel’s 15 years, but its choice is clearly not related to the AVX-512. Torvalds is like this, it attacks anyone, in no uncertain terms, to take or leave. Certainly in his invectives they raise issues of interest to the community, often leading to fruitful debates and even to directing the future choices of companies.