With the CPUFreq fix landing in Linux Git this week, the mainline Linux 5.11 kernel in its near final state looks very good for AMD Zen 2/3 hardware from Ryzen laptops and desktops via EPYC servers. The Linux 5.11 development kernel has been declining for most of the past two months, but now that frequency invariance regression has been addressed, not only has the regression gone, but it generally performs much better when compared to previous kernel versions.
Continuing in the series of benchmarks since December when I first encountered the regression on Zen 2 / Zen 3 when using the standard Schedutil regulator, here are my final benchmarks looking at AMD Linux 5.11 performance now the solution through of the said CPUFreq change landed this week and Linux 5.11 stable should come out on Sunday.
On the desktop side, I ran some fresh Linux kernel benchmarks with the Ryzen 9 5900X desktop at standard speeds, ASUS ROG CROSSHAIR VIII HERO motherboard, 2 x 8GB DDR4-3600 memory, 1TB Samsung 980 PRO NVMe SSD and Radeon RX 5600 XT- graphics. From installing Ubuntu 20.10, I tested Linux 5.9.16, Linux 5.10.12 and Linux 5.11 Git as the three latest stable kernel series at the time. The Ubuntu Mainline Kernel PPA was used to retrieve these vanilla kernel builds in an easily reproducible way. The default CPUFreq control on the recent kernels for AMD hardware is Schedutil.
On the pages that track the prominent individual benchmarks, let’s immediately take a look at the geometric mean of the 63 benchmarks that ran this round.
Linux 5.9 / 5.10 performed roughly the same overall, while with this actual Linux 5.11 end state, the performance has increased by 5.8% across the series of 63 different test cases. Not bad as the performance was lower for the past two months due to the regression, but luckily everything is now in working order if you use the standard Schedutil governor (as shown in previous tests, the ondemand governor is also better off with Linux 5.11) .
The Linux 5.11 kernel came first on this Ryen 9 5900X box for 60% of the benchmark. In the cases where 5.9 / 5.10 won, it was often with small margins.
Or last, the Linux 5.11 Git kernel was in last place for only 5 of the 63 benchmarks. So things are looking pretty good at this point with Linux 5.11 providing a better performing experience than previous kernels …
I also ran new benchmarks on the AMD EPYC 7F72 2P server with Supermicro H11Dsi-NT v2.0 motherboard, 16 x 8GB DDR4-3200 memory and 1TB Western Digital BLACK SN850 NVMe SSD. Ubuntu 20.10 ran on this AMD EPYC Zen 2 server while the Linux 5.9.16, 5.10.13 and 5.11 Git kernels were used last when starting those tests.
42 tests were performed on this EPYC server in the three latest kernel series, with more emphasis on server / workstation load. With this set of benchmarks, Linux 5.11, based on the geometric mean of all benchmarks, came out 6.0% higher … So quite similar to the 5.8% with the 5900X, even with the difference in tests. Meanwhile, it was the leader in 76% of the benchmarks compared to 5.10 / 5.9 and was in the last position for only 11% of the tests.
Let’s look in more detail at the Ryzen 9 5900X and EPYC 7F72 2P performance with Linux 5.11, which should be the version that powers Ubuntu 21.04 this spring.