How To Use The New AMD P-State Driver With Linux 5.17

[middy]

As said above, the benefit is more for laptops, but it might help your desktop save power without hurting performance. Also, some mobos will not have the necessary CPPC support enabled.

why does cppc provide a solid performance improvement on windows compared to linux? does windows actually have a better scheduler now?

[sagbobbit]

Is there a good chance Phoronix will be conducting benchmarks of this?

[spstarr]

You may also just add rd.driver.blacklist=acpi-cpufreq modprobe.blacklist=acpi-cpufreq in grub config to disable it as well.

From AMD folks,

"The amd_pstate module is not started at the boot because it's new module for the OS. You can modify the /etc/modules-load.d/modules.conf to add one line "amd_pstate". And then update-initramfs, reboot. Or you can blacklist acpi-cpufreq module and modprobe amd_pstate manually".

So all methods seem to work as long as acpi-cpufreq is not loaded or initcall_blacklist=acpi_cpufreq_init

[Soul_keeper]

I got the amd-pstate setup and working on 5.17rc8 over the weekend.
It works, reads the cppc data, but the preferred core "abstract scale" numbers are not actually used by the scheduler it seems.
Out of curiousity I modified the kernel drivers/acpi/cppc_acpi.c to use custom cppc data, no change in performance or core loading behavior
This imo would be the main benefit of the amd-pstate, being able to use and have a bias for the most efficient cores (ones with lower voltages for certain clocks). Lightly threaded workloads relying on boost will use fewer watts. This will leave more ppt budget for other things in thermally limited situations.

Also, it seems to me that the CPPC data is useless once someone enables custom PBO settings and sets up a custom curve offset in the bios. The data is "fused" on the cpu and static.

[ntropy]

Seems to be working great on my 4800H. I am at a good 6.5 W idling but that is not very different from what it was before. Yet when working in Unity with not too much GPU load, I get another hour battery time out the new scaling driver, to a good 5 hours @ 45Wh. https://i.imgur.com/w61JzjF.png Laptop feels way cooler too.

Powertop seems to be a bit all over the place now: https://i.imgur.com/qWLldsd.png

[Soul_keeper]

I just modified the cppc_acpi.c to use Intel's ITMI with ryzen ... performance goes up slightly (~5%).
I can see it loading my most undervolted (curve offset), cores now. Using a custom cppc high for each core also.
So I guess this confirms that preferred cores is still not implemented with the scheduler yet. Atleast the ground work is in place.

[Linuxxx]

Michael

Since when did Ubuntu's "generic" kernel flavor switched over to activate full "PREEMPT" by default?

Never seen that before...

Can You pinpoint when that intrusive change to Ubuntu's default kernel config happened?

[birdie]

I'm not using it for my Zen 3 system yet.

[Linuxxx]

I'm not using it for my Zen 3 system yet.

Good to see You are still alive & kicking, Artem!

[birdie]

Good to see You are still alive & kicking, Artem!

Alive, yes, kicking, nope. Thanks for caring anyways.