GCC 12 Compiler Squaring Away Its AVX-512 FP16 Support

[coder]

Sure. I think for Intel it is just a matter of market differentiation.

I doubt that. Intel has rolled out AVX-512 in 2 generations of laptop CPUs (Ice Lake and Tiger Lake) and now one generation of desktop CPUs (Rocket Lake). It was surely painful for them to walk that back. And if they want good adoption of AVX-512 in software for servers & workstations, they really need to get it into the hands of the typical developer.

I think they blithely thought they could manage AVX-512 threads at the OS level, using SIGILL exceptions to fault AVX-512 threads over from little cores to big ones. Then, when this turned out to work poorly in practice, then they decided to disable it on the big cores. Time will tell, but I think when people get Alder Lake under a microscope, they'll see the AVX-512 logic in the big cores, sitting dark.

Intel can't walk away from AVX-512, in their mainstream products. I think we'll see it come back, once they add it into their little cores, which (starting with Alder Lake) will finally now have 256-bit AVX/AVX2.

[Paradigm Shifter]

Handling of AVX-512 by Intel has been a mess from day one. The sub-segmentation of different instructions in the "instruction set" means that an AVX-512 capable CPU does not have the same capabilities as another AVX-512 capable CPU. The issues with die shrinks I'm sure also caused issues, which slowed adoption, which means that no one coded for it except in odd circumstances... yeah, it was a pain.

I've seen reports from Tom's Hardware (make of that what you will) that Zen 4 will have AVX-512... likely too late to walk that back now. The real tell would be if Zen 5 did. Or whether Intel put it back in whatever CPU is after Alder Lake.

AMD could perhaps "save" an Intel instruction set...

But it'll all be down to compiler. Even enabling the AVX-512 flags, using the Intel compiler offers a huge performance boost over gcc (9/10/11) in the AVX-512 application I use.