Anon (no.delete@this.spam.com) on July 21, 2022 1:25 am wrote:
> Wilco (wilco.dijkstra.delete@this.ntlworld.com) on July 19, 2022 6:15 pm wrote:
> > Where do you get that 0.15MB from?
>
> Altra 80 32MB/80 cores = 0.4MB per core
> Altra 128 16MB/80 cores = 0.25MB per core
>
> > The MCF results show that for 128 threads a
> > 128MB or 256MB L3 is much better than a 16MB L3. That's a 8x to 16x increase.
>
> You are comparing different cores, there is no same core, same
> compiler flags, etc comparison between 16MB and 128MB L3.
>
> > What makes you think it matters?
>
> Anadtech Epyc 7763 MCF result: 158
> Official Epyc 7763 MCF result: 464
>
> Yes, it matters.

They used a cheating compiler for the official result. Different compilers so impossible to
conclude anything about flags or allocator.

> > 1MB per core absolutely makes a huge difference as the results show. Of course
> > if you already have 4MB per core then tripling does not help much.
>
> No! What the comparison between Yitian and Anandtech Altra Max shows
> is that the bellow compiler flags makes a huge difference:
>
> -mabi=lp64 -std=c99 -L/opt/rh/devtoolset-10/root/usr/bin//lib64 -L/opt/rh/devtoolset-10/root/usr/bin//lib
> -L/home/jemalloc-5.2.1/lib -g -O3 -mcpu=neoverse-n1 -funroll-loops -flto=32
> --param early-inlining-insns=96 --param max-inline-insns-auto=64 --param inline-unit-growth=96
> -fgnu89-inline -ljemalloc

This proves nothing: these are different SoC so I fail to see how you can attribute the
difference to compiler flags or allocator. Add to that you are comparing llvm vs gcc.