The Linux Performance For AMD Rome vs. Intel Cascade Lake One Year After Launch

[phoronix]

Phoronix: The Linux Performance For AMD Rome vs. Intel Cascade Lake One Year After Launch

With the Intel 2nd Gen Xeon Scalable (Cascade Lake) processors having turned a year old in April and next month marking one year since the launch of the AMD EPYC 7002 (Rome) series, here are fresh benchmarks of the dual Xeon Platinum 8280 versus the AMD EPYC 7742 when testing the Linux software stack from early 2019 and then again using a bleeding-edge Linux software stack as of this month. This shows how the Linux software performance has evolved over the past year for both Intel and AMD on the server front as well as how the current top-end SKUs are competing right now.

http://www.phoronix.com/vr.php?view=29402

[edwaleni]

Intel Java performance really improved with the later OS updates.

[CochainComplex]

I would like to know what the unit of measurement used in the final "Geometric Mean Of All Test Results" chart is.

man didn't you here the intel people? benchmark is not important so just take this numbers as feelings
300 feels better then 280

[caligula]

man didn't you here the intel people? benchmark is not important so just take this numbers as feelings
300 feels better then 280

Intel is perfect if you need RHEL style stability. Last time I heard, 14nm process could be abandoned in 2035. However, each 14nm+++ minor upgrade requires a new 100% incompatible socket and a brand new $250 motherboard.

[lowflyer]

Michael, thanks again for this interesting test series.

While looking through all these graphs, I had the impression that the new OS sort of "leveled the playing field". Both processors profited significantly. There were the usual ups and downs. But I have the impression that both processors are "more equal" than with the old OS. The geometric mean is an interesting figure if you want to see the "final performance". But how these processors moved "closer to each other" is not visible there. Would it be possible to show a graph with something like "results distribution"? I think this could add quite some value to these graphs.

[Citan]

Hi Michael thanks for this new series of tests.

I have a small wish that you may maybe grant: we've seen all these past years how Intel "cheated" on performance by using dangerous shortcuts, and how all the palliatives that have to be applied took cumulative toll on performances.

I seem to have some memory that you planned (or already did) updated comparative of processor performances with all those applied.
If you still have an old-gen AMD (like a Phenom II) and similar-gen Intel competitor, could you please make a batch of tests to see how those oldies would fare in nowadays context (security patches on intel procs, apps being overall much more multithreaded than before)?

So we can see, among other things, how much betting on multicore instead of high frequency when investing in 2010-2012 payed off in the long run.
Thanks in advance.

[jbrnd]

Was the Intel MKL library installed anywhere on the system? If so, it's possible that this gave the Intel system an advantage on the tests where a BLAS library was used. In my limited testing on an AMD CPU, MKL was three times slower than OpenBLAS for eigenvector operations (the difference was smaller for other operations).