AmpereOne Performance Scaling From 32 To 192 Cores, Core-For-Core Benchmarks Against Ampere Altra Max

[phoronix]

Phoronix: AmpereOne Performance Scaling From 32 To 192 Cores, Core-For-Core Benchmarks Against Ampere Altra Max

Earlier this week I began with AmpereOne A192-32X benchmarks and will continue for the next several weeks in finally having hands-on with the 192-core AArch64 server processor using a Supermicro ARS-211M-NR R13SPD 2U server platform. In today's next phase of AmpereOne performance benchmarking is looking at how AmpereOne scales across 32, 64, 96, 128, 160, and 192 core counts plus seeing core-for-core at 128 cores how AmpereOne compares to the Ampere Altra Max M128-30 processor. Plus these AmpereOne benchmarks at varying core counts against the AMD EPYC and Intel Xeon competition.

https://www.phoronix.com/review/ampereone-benchmarks

[raystriker]

If I'm interpreting things correctly, this bodes seriously well for AMD's Turin (Zen 5C Epyc). Should gap the Ampere One comfortably

[JanW]

The Ampere One somehow is now beating the EPYC 9684X (1P) slightly in geo mean, whereas it was trailing by 15-20%, depending on benchmark selection, in the earlier review. Is this just a more favorable selection of benchmarks this time around? Or am I reading this wrong?

[Michael]

The Ampere One somehow is now beating the EPYC 9684X (1P) slightly in geo mean, whereas it was trailing by 15-20%, depending on benchmark selection, in the earlier review. Is this just a more favorable selection of benchmarks this time around? Or am I reading this wrong?

Largely the same tests but limited to those that are known to be very scalable to high core counts versus that and then a mix of other server workloads that don't always scale so well but still relevant.

[sophisticles]

Scaling benchmarks really say more about the way the code is written than about anything else.

As shown, some tasks scale well with thread count and some don't and the CPU used is not going to change that all that much,

[purpleduggy]

ARM and RISCV is coming. X86 is done.

[F.Ultra]

Scaling benchmarks really say more about the way the code is written than about anything else.

As shown, some tasks scale well with thread count and some don't and the CPU used is not going to change that all that much,

much more important is if the data is shared or not. When you have to perform operations on the same data set then scalability tanks due to cache trashing, that is why the applications in this benhmark that can handle completely different data sets on each core scales so much better than the ones that have to share the data.

[coder]

Scaling benchmarks really say more about the way the code is written than about anything else.

In either case, scaling analysis generally tells you something about the hardware, so long as the benchmark is remotely scalable (and here's where GeekBench 6 MT sucks). If a lot of synchronization and data-sharing is happening between threads, then such analysis would show you how well different hardware handles that stuff (assuming he did comparative scaling analysis on different machines, which he didn't).

If the benchmark is highly-scalable, then it absolutely tells you about the hardware. The tricky part is that you need to have some a priori knowledge of how well the benchmark theoretically can scale, which is lacking here.

SPEC2017_rate is one of the best benchmark suites for testing scalability on shared-nothing workloads, since it involves running N copies of each benchmark. Sadly, few hardware reviewers are still running it, these days.

[coder]

ARM and RISCV is coming. X86 is done.

EPYC 9754 beat it on perf/W by 1.8%.

Turin is coming later this year. If this leak is true and other workloads have similar improvements, it's going to be a very bad day at Ampere, when it launches:

https://wccftech.com/amd-epyc-9755-turin-zen-5-cpu-128-cores-twice-as-fast-epyc-9754-bergamo-zen-4c/

[ddriver]

ARM and RISCV is coming. X86 is done.

Do yourself a favor and stop being dumb already. uArch is irrelevant, the quality of chip design is all that matters.

What did x86 did to you anyway? Touched you in a bad place or something?

Besides, x86 is more or less dead already, we are using x86-64 now. And by the looks of it, it ain't going nowhere.