AMD EPYC 7302 / 7402 / 7502 / 7742 Linux Performance Benchmarks

[tchiwam]

Do you understand why is that? My crystall ball says one of the things you want to look into is cache utilization. And you might want to pick the cpu configuration that gives you the largest caches.

That's why I am interested in the 7262 or the maximum of L3 for 8 cores or the 7302. I think I can fit my data in the closest L3 for reuse with 2 threads per tasks.
I do need ECC memory for the 3 months of crunching needed, ECC 3200 UDIMM is harder to find than the ECC RDIMM 3200... So my ram might make the EPYC cheaper than a TR...

[pegasus]

I have a load that is running better as many single threads.

Do you understand why is that? My crystall ball says one of the things you want to look into is cache utilization. And you might want to pick the cpu configuration that gives you the largest caches.

[Michael]

I have a load that is running better as many single threads. Am I better of with 2 cpu with higher base clock or 1 with an higher boost clock ? I think this is missing in the test suite.

New test contributions / scripts always welcome.

https://github.com/phoronix-test-suite/test-profiles/

Thanks.

[tchiwam]

I have a load that is running better as many single threads. Am I better of with 2 cpu with higher base clock or 1 with an higher boost clock ? I think this is missing in the test suite.

[Michael]

Great benchmarks! Any chance you could take a look how they do with running python+numpy in MKL vs. OpenBLAS against the Intel CPUs? We would love to get the next servers for our scientific computing needs to be AMD, but so far the complexity of MKL issues always got us to choose the safer option.

If you can provide a reference script on how to setup and run any desired tests in an automated and reproducible manner, I'm always happy adding new test profiles to PTS.

[wonko]

Great benchmarks! Any chance you could take a look how they do with running python+numpy in MKL vs. OpenBLAS against the Intel CPUs? We would love to get the next servers for our scientific computing needs to be AMD, but so far the complexity of MKL issues always got us to choose the safer option.

[caligula]

I guess by now it's pc to say the Intel solutions are screaming "Help! I'm getting raped!" come the perf*dollar charts...

They'll just need to cut down the prices by 80%. Shouldn't be too hard.

[Mark Rose]

Interesting to see PHP do so well on the Xeons. How does Clear's optimized PHP perform on the newest Epyc and newest Xeons?

[grigi]

I see. That makes sense.

[pegasus]

What does a Epyc server cost without the cpu? What does a Xeon server cost without the cpu?

As long as you stay with modest amounts of memory, cost difference is negligible and cpu is the main cost factor. Once you get into multi-terabytes memory territory, then memory becomes your largest price chunk and cpu doesn't play that much of a role. Still, it's cheaper with epyc since it can drive 8ch memory with up to 2TB per socket.