NAS Parallel Benchmarks: EPYC 7601 vs. EPYC 7742 vs. Xeon Platinum 8280

[phoronix]

Phoronix: NAS Parallel Benchmarks: EPYC 7601 vs. EPYC 7742 vs. Xeon Platinum 8280

Not included as part of our original EPYC 7742 / EPYC 7002 "Rome" Linux benchmarks was the NAS Parallel Benchmarks (NPB) developed by NASA. While an MPI testing favorite, there were build issues with the older version of NPB packaged by the Phoronix Test Suite. But with recently having updated that test profile against the latest NPB upstream, here are some results for the EPYC 7742 2P, EPYC 7601 2P, and dual Xeon Platinum 8280 benchmark results. Separately, there's also results now for NeatBench 5 with this video editing plug-in test case now part of the Phoronix Test Suite...

http://www.phoronix.com/scan.php?page=news_item&px=NPB-EPYC-Rome-Benchmarks

[tomtomme]

something is wrong with the harmonic mean plot. It looks like the EP.D test, which is the only AMD lost. Or is this mean weighing the benchmarks maybe by duration?

[Michael]

something is wrong with the harmonic mean plot. It looks like the EP.D test, which is the only AMD lost. Or is this mean weighing the benchmarks maybe by duration?

https://openbenchmarking.org/result/...or=y&obr_rro=y There is more data in the OB result file. Basically some extra tests of NPB not included as they didn't add any extra value to this short article.

[grigi]

After using harmonic mean more and more, I found it is only ever accurate when you compare the EXACT same metric. When the metric has a slightly different meaning the results can be non-intuitive.
It is however the perfect mean to use to determine race-to-finish from a whole bunch of tasks that have to be finished (as in no allowance of "dropped" frames, e.g. FPS in games)(FPS in scene rendering is however correct use)

Generally the Geometric mean gives the best overall result.

[Michael]

After using harmonic mean more and more, I found it is only ever accurate when you compare the EXACT same metric. When the metric has a slightly different meaning the results can be non-intuitive.
It is however the perfect mean to use to determine race-to-finish from a whole bunch of tasks that have to be finished (as in no allowance of "dropped" frames, e.g. FPS in games)(FPS in scene rendering is however correct use)

Generally the Geometric mean gives the best overall result.

Right normally I use geometric mean. In this case for the NPB tests, it is the same exact metric of Mop/s.

[M@GOid]

The server market is the cash cow for Intel, with huge margins and total market dominance. Test after test it becomes harder to justify buying anything from them since Epic. I can only imagine what Intel salesman have to do to convince a client and justify the high prices of their offerings. A client's workload must really favor Intel's architecture to make ends meet.

[pegasus]

Intel salesdrones are offering you vnni instead of nvidia gpus; avx512 instead of epyc; whole software ecosystem (compiler, mkl, many other tuned libs, ...) instead of "that opensource crap". There are still people out there who buy into this. What I find suprising is that Intel is not yet ready to offer massive discounts on their cpus. Or maybe it is and we weren't buying enough to qualify ... But eventually they will have to start price-compete with AMD.

[tomtomme]

https://openbenchmarking.org/result/...or=y&obr_rro=y There is more data in the OB result file. Basically some extra tests of NPB not included as they didn't add any extra value to this short article.

thanks for clearing that up