David Kanter (dkanter.delete@this.realworldtech.com) on July 12, 2020 5:01 pm wrote:
> Linus Torvalds (torvalds.delete@this.linux-foundation.org) on July 11, 2020 11:41 am wrote:
>
>
> > I've said this before, and I'll say it again: in the heyday of x86, when Intel was laughing all the
> > way to the bank and killing all their competition, absolutely everybody else did better than Intel
> > on FP loads. Intel's FP performance sucked (relatively speaking), and it matter not one iota.
>
> I totally agree. Integer workloads are the most important.
>
> > Because absolutely nobody cares outside of benchmarks.
>
> I realize you are being your usual hyperbolic self, but I don't
> think that's true. There are plenty of workloads that use FP:
>
> 1. HPC
> 2. Machine learning

But are those key workloads for the vast majority of their customers?
Probably not a majority of server, and certainly not for client. You could argue about what the balance of server vs client is for Intel, but regardless maybe we'd all be better served if they had a clear "FP heavy workloads" core vs. "everything else" core. Obviously, Intel did have this split with Xeon Phi, and decided to go the opposite direction by folding AVX512 into mainstream server, aka all of this convergence-optimized core the marketing keeps parroting.

From the outside, it seems more like a late reaction to the preeminence of GPUs in HPC and ML, and Intel not wanting to acknowledge it until very recently.

>
> > The same is largely true of AVX512 now - and in the future. Yes, you can find
> > things that care. No, those things don't sell machines in the big picture.
> >
> > 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.
>
> Not all workloads that want FP compute actually run on the GPU. For example, many
> neural networks are trained on CPU because of the required memory capacity.

Maybe that's more an issue with GPU memory capacity, and the GPU programming model in general. Supercomputers seem to have decided this programming model is an issue, since all 3 of the announced exascale machines are going for some sort of shared/coherent memory model between CPU and GPU.

>
> > 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 once calculated how much area the AVX512 units used on SKL-SP. It's a fair bit on a per-core basis
> (around 1mm?). But that would only let you add 1-2 cores (because each core needs L3 cache, etc.).
>
> Also, with Icelake, there should be much less of a frequency hit because it has a vastly better DVFS system.
>
> > Yes, yes, I'm biased. I absolutely destest FP benchmarks, and I realize other people care deeply.
> > I just think AVX512 is exactly the wrong thing to do. It's a pet peeve of mine. It's a prime example
> > of something Intel has done wrong, partly by just increasing the fragmentation of the market.
>
> Yes the market fragmentation is stupid.
>
> David