Linus Torvalds (torvalds.delete@this.linux-foundation.org) on May 21, 2022 4:58 pm wrote:
> Of course, maybe Intel can make a small and "good enough" AVX512 unit so that this
> becomes a non-issue for that reason - support it on every CPU, just in a weaker
> format (but not so weak that using AVX512 is slower than the alternative).
>
> I just personally suspect that AVX512 was never designed for that, and
> just doesn't scale down very well. But I'm not a hardware engineer.

All current Intel P cores do support it. AMD's upcoming Zen4 will support it, and they've suggested their smaller Zen4c cores will as well.
Even CNS supports it, and it's a reasonably small x86 core (considering the node it's on), and it performs well:



In other words, all x86 cores will support it, with decent perf; the exception currently being Intel's Atom line, which never supported any AVX in the past (ignoring the Xeon Phi line), and only got AVX2 in the latest release (which could be seen as a stepping stone towards AVX512, but that's just pure speculation).

There doesn't seem to be much to suggest that AVX512 requires much die space - Intel's early approach (server) was quite aggressive, but AMD/Centaur is less so.
The only remaining question is whether Intel can get it into their E core lineup - if so, it'll meet your "support it on every [new] CPU" requirement (and won't be slower than AVX2). If not, support for it on Intel consumer platforms will likely be dead, though with AMD gaining marketshare there, it'll likely hang around.