none (none.delete@this.none.com) on August 14, 2018 5:15 am wrote:
> Jeff S. (fakity.delete@this.fake.com) on August 14, 2018 5:06 am wrote:
> > none (none.delete@this.none.com) on August 14, 2018 1:43 am wrote:
> > > y-cruncher is very peculiar in its heavy use of AVX2 and AVX-512. For instance it got a
> > > speedup of 20-30% with AVX-512.
> >
> > There were multiple things that seemed off to me about Ian's review (mostly around the totally
> > unexplained power measurements and analysis), but the y-cruncher results don't look weird to
> > me whatsoever. Alex, the author of y-cruncher and coincidentally also my colleague, also knows
> > Ian I think, so I may have a back channel in getting parts of the review clarified.
>
> Just to clarify, I didn't say the y-cruncher results are wrong or demonstrate a poor software
> implementation. It's just that it seems heavily tuned to benefit the most from
> multiple FMA (FFT-based multiplication) and the author is very good at that :-)
>
> So what this test shows is that Intel cores are much better at heavily tuned FP (FMA)
> than AMD cores.

Y-cruncher's algorithm is surprisingly bandwidth dependent across all orders of size magnitude, so as much of the improvement may come from Skylake's wider read datapaths from L1D and L2 as anything, but clearly Zen (1/1+) is not meant to compete with Intel in serious AVX work.

My comment stems from the fact that a lot of the WIP review's analysis seems to draw numbers from thin air so far, and I thought you may have possible insinuating that it might have some other elements of Intel-friendly benchmark choices baked in.