Michael S (already5chosen.delete@this.yahoo.com) on July 11, 2020 12:58 pm wrote:
> Linus Torvalds (torvalds.delete@this.linux-foundation.org) on July 11, 2020 11:41 am wrote:
> >
> > And AVX512 has real downsides. I'd much rather see that transistor budget used on other things that
> > are much more relevant. Even if it's still FP math (in the GPU, rather than AVX512). Or just give me
> > more cores (with good single-thread performance, but without the garbage like AVX512) like AMD did.
> >
> > I want my power limits to be reached with regular integer code, not with some AVX512
> > power virus that takes away top frequency (because people ended up using it for memcpy!)
> > and takes away cores (because those useless garbage units take up space).
> >
>
> I do believe that AVX512 have significant area and power cost in SKX/CascadeLake,
> even in bastardized Gold5xxx/Silver/Bronze variant.
> But is it still true in client variant of Sunny Cove?
> What silicon experts would say?
>

Well there are those die shots from Fritz where you can directly compare SKL and SKX.

If you assume that SKL is a relatively optimal design for AVX2-only, then the area cost is low single digit % in the core, *if* you compare against 1 FMA and ignore the extra 768 KiB of L2 cache. The SKX core and the SKL core have the same dimensions, and the only significant parts of any size that are missing (masked out) in SKL are the register file and something taking maybe 20% of the vector EUs (probably the AVX-512 shuffles), as well as some structure in the load-store unit that has doubled in size (maybe the store buffer).

Of course, I think this "SKL is a relatively optimal design for AVX2-only" is probably totally false. The fact that they are almost exactly the same core, with those tweaks means that SKL was probably either designed as 512-bit capable and then had that capability removed at some point, or was designed as 256-bit but always meant to share almost all of the floorplan with the later 512-bit part.

So we don't really have a direct comparison. Maybe 14nm Broadwell would be the best comparison: what are the sizes of Broadwell vs Skylake core? After a quick search I couldn't find the relative core sizes.

Disclaimer: not a silicon expert.