Anon (no.delete@this.spam.com) on July 17, 2022 8:46 pm wrote:
> Wilco (wilco.dijkstra.delete@this.ntlworld.com) on July 17, 2022 3:50 pm wrote:
> > Having a lower frequency and hardly any L3 cache. If you look at the subscores, you
> > can tell that it does well on anything that fits in L2, but actually scores far worse
> > than Altra in memory intensive benchmarks. So is your conclusion that the memory bandwidth
> > was reduced in Altra Max? Or is it simply lacking a properly sized L3?
>
> So, you think 0.4MB L3 per core vs 0.25 MB L3 per core made a huge difference?
>
> Let's me give you another hypothesis: all subsets Altra Max performed worse was
> because it was bandwidth starved and, as usual, when you add threads to an already
> bandwidth starved workload you reduce performance by worsening locality.
>
> If you disagree, how you explain so a little (really litle, 0.15 MB per core
> with each core already having 1MB L2) L3 giving such performance delta?

Ever heard of tipping points or disk thrashing? If you understand that you should be able to get that the same happens when your working set does not fit in your caches, resulting in memory latency going through the roof. This has absolutely nothing to do with bandwidth.

Altra Max achieves far higher bandwidth than EPYC for any number of cores. EPYC is ~2.3 times faster on MCF despite the far lower bandwidth. And Yitian is ~3.6 faster than Altra Max on MCF with a 50% increase in memory bandwidth.

Could it be that big L3 caches actually help, or do you think AMD was stupid to add a huge L3 to their design?

Wilco