Ryzen 7 2700X CPUFreq Scaling Governor Benchmarks On Ubuntu Linux

[phoronix]

Phoronix: Ryzen 7 2700X CPUFreq Scaling Governor Benchmarks On Ubuntu Linux

With this week's Ryzen 5 2600X + Ryzen 7 2700X benchmarks some thought the CPUFreq scaling driver or rather its governors may have been limiting the performance of these Zen+ CPUs, so I ran some additional benchmarks this weekend...

http://www.phoronix.com/scan.php?page=news_item&px=Ryzen-7-2700X-Scaling-CPUFreq

[R41N3R]

Based on these results it seems that gamemode (and by this the performance governor) is not required, when using schedutil with a AMD chip. Schedutil is the default on Arch Linux as far as I know for AMD.

[dungeon]

I think we commented about performance of ditching governor altogether and not testing its levels

Fastest of these all is of course no governor at all. Further i guessed that on ubuntu kernel specifically that can't be easely disabled/blacklisted since acpi_cpufreq is not built as module

[monte84]

A concern I was raising is if the Ryzen CPU was hitting it's turbo clock speeds. On my 1700 when setting performance governors I cannot achieve turbo speeds and am left at the base clock frequency. This obviously won't affect multi core speed( although the 2700X is supposed to be better about turbo clocks over more cores. And also Intel MCE (multi core enhancement) can clock all cores to it's turbo speed, assuming power and temperature requirements are met. Can you watch CPU clock speeds? With the different governors to see if turbo is being used properly? watch grep "cpu MHz" /proc/cpuinfo

[leipero]

Something interesting for AMD/Intel CPU designers to consider. Obviously I have no clue what I'm talking about, but, it seems to me that it is possible to add functionality of frequency and voltage range in CPU/UEFI that is completely independent of OS, so, CPU detects "activity" and clock up higher. If i understand this corectly, that is actually how modern CPU's work, with one difference that instead of CPU "detecting activity" OS/kernel do send that information to UEFI that goes to the CPU, therefore making latency and potential problems, not because CPU system is bad, but simply because steps taken in the process.

Ditching that "activity detection" completely from the OS level, and moving it preferably to the CPU would solve such problem, and, in case someone doesn't want those features, he could disable it in UEFI or even with option avialable in the OS itself.

I remeber when moving system from Windows to GNU/Linux, I had "wow" moment in browsing experience, because Windows 7 didn't have proper acceleration it couldn't follow refresh rate etc., plus browsers were in worse state. But I've faced the problem with Cool and Quiet, C6 states etc., so I've disabled everything in UEFI, downvolted CPU and solved problem on hardware level. That is the way to solve problem IMO, and if done properly, no one would even need to bother with software and disable anything.

[numacross]

For Intel what you're describing leipero is Speed Shift (https://www.anandtech.com/show/10959...00k-i3-7350k/3) and for AMD I'm pretty sure that's how XFR(2)/Precision Boost works as well.

[leipero]

numacross Hm, I thought that XFR(inncluding ver 2) is new "turbo" method. Either way, I can't speak for Intel, but from the test it seems that for both Intel and AMD they still depend on OS, because if system described above is fully working, OS should have nothing to do with it, aside from potential off/on function (as it is in UEFI). In this system, CPU itself would detect activity completely independent of OS, and user would have an option just to enable or disable (and potentially tune) that feature.

[monte84]

numacross Hm, I thought that XFR(inncluding ver 2) is new "turbo" method. Either way, I can't speak for Intel, but from the test it seems that for both Intel and AMD they still depend on OS, because if system described above is fully working, OS should have nothing to do with it, aside from potential off/on function (as it is in UEFI). In this system, CPU itself would detect activity completely independent of OS, and user would have an option just to enable or disable (and potentially tune) that feature.

XFR is beyond normal turbo clocks, XFR (eXtended Frequency Range) with thermal and power consumption that can be restraints on XFR.

[leipero]

monte84 Yeah but it doesn't regulate power states?

[Zan Lynx]

As far as I know about it, the only thing that CPU governors can do to Ryzen is slow it down. If you put it in performance and leave it alone, it should use XFR and manage its own power.