Linux's P-State Performance Governor Shows Unexpectedly Big Boosts For The Intel Core i9-11900K

[andre30correia]

I use always the cpufreq driver, this pstate draws a lot of more power, burns the cpu even in idle never understand why, I want my desktop and laptop cool if I need more response I push to performance

[hoohoo]

Re: power consumption, this is just exploiting the full power dissipation capacity of Intel.

[Linuxxx]

Michael
I feel like throwing the »schedutil« CPU governor into the mix would lead to some interesting insights, too; particularly on Rocket Lake's one-of-a-kind CPU cores. (Maybe "little Rocket-Man" from North Korea would be interested in actually buying them?)

And should really You plan to do these benchmarks (which would be great, BTW), testing both possible CPU drivers on Intel (namely intel_cpufreq vs. acpi-cpufreq) would make them even more valuable!

You can activate the first one by booting the Linux kernel with the following parameter:

intel_pstate=passive

As for the second option:

intel_pstate=disable

I believe such benchmarks have never been conducted before, where "schedutil" dukes it out with itself on the same CPU but on differing drivers, which would mark a world-premiere!

[arabek]

I wonder why aren't you testing against the schedutil gov? It should yield better thermal/power consumption to performance ratio than the powersave one, and not be as power-hungry as the performance one.

[MadCatX]

I wonder why aren't you testing against the schedutil gov? It should yield better thermal/power consumption to performance ratio than the powersave one, and not be as power-hungry as the performance one.

Intel chips have been doing performance scaling in hardware since SKL. intel_pstate only hints them about the target energy consumption. Older Intel chips which don't have HWP rely on intel_pstate scaling governors. In this case, intel_pstate's performance is like the acpi-cpufreq's performance and intel_pstate's powersave is similar to ondemand and schedutil. Unlike acpi-cpufreq, intel_pstate governors read data from specific registers on the CPU to optimize P-State selection and should do a bit better job that the generic acpi-cpufreq governors because of this.

[gnartz]

Michael : Just came across your article after trying to enable pstate on 11900K with 5.12.0-051200rc8-generic_5.12.0-051200rc8.202104182230_amd64 but I can't seem to get it running. Dmesg gives

intel_pstate: CPU model not supported

- any ideas?

Thanks.
g.

[rsr911]

Latest Asus bios for my Rog Strix Z590 E solved this issue for me. On powersave I can go from 800-5.4 and run at 5.1 under heavy load. Still amazed the authors Noctua air cooler is handling 380+ watts when my 360 AIO thermal throttles around 320 watts with a lapped water block and CPU and liquid metal interface.

[gnartz]

Thanks! Unfortunately, despite running the latest BIOS on my Z590 Apex, I'm getting the same error message.

May I ask: What determines whether you get 54 or 51? Is that your AVX offset? Or does per-core clocks work for you under Linux? That is, setting 54/53/52/... in BIOS to get different max clocks depending on thread load? In my experience, since the load is distributed across cores, no matter what per-core settings I apply, the CPU will always run at the lowest of them.

[Eumaios]

Goodness: 30% higher average power consumption, 20% higher average CPU temperature--and 10% higher average performance. I suppose it's fine if you're on a desktop with good cooling, but on a laptop . . .