Linus Torvalds He does not mince words and often speaks loud and clear about what he thinks of some things. This time he has decided to say what he thinks about the Intel AVX-512 instruction set. An instruction set that is added on top of the base x86-64 ISA to provide specially designed calculation instructions for HPC environments.
In principle these instructions should be good, in fact they add performance for certain types of calculation applications used in HPC environments. But Linus Torvalds did not like that this set of instructions is also included in the Intel desktop processors. In fact, it seems that Intel has reconsidered and it seems that their futures Alder Lake will not have AVX-512 support.
Linus Torvalds is of the opinion that they should focus more on what really matters for this type of segment rather than wasting resources on new instruction sets like AVX-512 that are meaningless. out of market HPC (High Performance Computing).
Since AVX-512 debuted in the Xeon Phi x200 (Knights Landing), then it would move on to Skylake-SP, Skylake-X, Cannon Lake, and Cascade Lake. After that, some like Cooper Lake and Ice Lake also supported certain subsets of instructions from the AVX-512 repertoire.
La opinion posted by Phoronix of Linus Torvalds it has no waste:
I hope that AVX512 dies a painful death, and that Intel starts to fix real problems instead of trying to create magic instructions and then create benchmarks that they can look good on.
Hope Intel gets back to basics - get your process working again and focus more on regular code that isn't HPC or some other nonsensical special case.
I've said this before, and I'll say it again: in the heyday of x86, when Intel was killing all the competition, absolutely everyone else did better than Intel at FP (Float-Point) loads. Intel's FP performance sucked (relatively speaking), and it didn't matter one iota.
Because absolutely nobody cared outside of the benchmarks.
The same is largely the case with AVX512 now, and in the future. Yes, you can find things that matter to you. No, those things don't sell machines in the big picture.
And AVX512 has real downsides. I'd rather see the transistor budget being used on other things that are much more relevant. Even if it is still FP math (on the GPU, instead of AVX512). Or just give me more cores (with good single-threaded performance, but without the garbage like AVX512) like AMD did. (Remember that now Linus uses AMD Threadripper on his PC)
I want my power limits to be reached with a regular integer code, not with an AVX512 power virus that removes the maximum frequency. Since that useless junk takes up space and removes cores.
If I realize. I absolutely study FP benchmarks and realize that other people deeply care. I just think AVX512 is exactly the wrong thing to do. It's a hobby of mine. It's a prime example of something Intel has done wrong, in part by increasing market fragmentation.
[…] Make a FPU that is good enough, and people will be happy. AVX2 is much more than enough.
Yes, I also agree with Linus. Intel has always shown something slightly good. Perhaps because it leans more towards marketing. But if you were really interested, you would have already included nano technology, with exactly more cores. Anyway, all this will only cause new companies to emerge doing what Intel does not want to do.
Intel's with their proprietary instruction sets are not very Intel-igent.