Linux 6.6's cpupower Utility Enables New AMD P-State Features

[phoronix]

Phoronix: Linux 6.6's cpupower Utility Enables New AMD P-State Features

Linux's cpupower utility lives within the Linux kernel source tree for reading and tuning various CPU power settings rather than poking at sysfs files directly or other means of adjusting your processor power-related tunables. With the upcoming Linux 6.6 kernel cycle the cpupower utility is adding support for adjusting new AMD P-State driver features...

https://www.phoronix.com/news/Linux-6.6-cpupower

[Namelesswonder]

I did not very scientific synthetic tests with all the different modes including the new guided mode to see how they would perform and my results pretty much lined up with the previous EPYC tests. I only tested with the performance governor since my 5900X gets constrained by the others. I tracked the end result and monitored the frequency with ryzen_smu and ryzen_monitor.

Guided autonomous mode came out on top, followed by active EPP mode, then with passive mode and CPUFreq being very close to each other. The difference was down to about 10-15MHz from guided to CPUFreq.

The differences do manifest themselves statistically in synthetic benchmarks, but with only a 0.6% difference at most in one test score and a 50 millisecond difference in another timed benchmark, it won't translate to anything meaningful in real world performance at least on desktops.

It will be interesting to see tests with guided mode on laptops and mini PCs. The power and thermal constraints led to a more meaningful impact with performance-balanced EPP on the laptops, so maybe guided mode might eek out a bit more.

[Linuxxx]

For benefiting both AMD and Intel P-State drivers, there is Energy Performance Preference (EPP) value change support. This allows adjusting the EPP value from the cpupower utility for indicating your energy/performance preference that in turn ultimately influences the CPU frequency scaling behavior. This is exposed via a new "epp" option for cpupower.

Intel CPUs are accepting values in the range between 0-15, whereas last I checked on AMD it was 0-255 for setting the EPP.

Is that still the case and if yes, why can't the developers settle on standard values across vendors for this?

[edwaleni]

Not all AMD CPU's support all the p-states. I don't have an "exact" accounting, but most 5000 series Ryzen desktop CPU's only support 3 states. It appears it increases with each new increment of Zen that is released.

[dougg]

Building tools supplied by the Linux kernel is a crap shoot. cpupower has a handcrafted Makefile that builds a shared library and puts it in /usr/lib64 . That looks okay for RedHat based systems, not so much for Ubuntu/Debian users. Perhaps RedHat and other companies that profit from the Linux kernel could contribute a few of their professional maintainers to clean up this sort of stuff.

[yump]

Not all AMD CPU's support all the p-states. I don't have an "exact" accounting, but most 5000 series Ryzen desktop CPU's only support 3 states. It appears it increases with each new increment of Zen that is released.

I don't have any of the relevant chips, but are you sure you're using amd-pstate? As I recall a very small number of states is what you see with the acpi-cpufreq driver.

[skeevy420]

Not all AMD CPU's support all the p-states. I don't have an "exact" accounting, but most 5000 series Ryzen desktop CPU's only support 3 states. It appears it increases with each new increment of Zen that is released.

3 states is an ACPI Ryzen thing. AFAIK, all Ryzens are like this (at least Zen 2 and up on AM4):

With acpi_cpufreq you'll get 3 states. Lowest, Middle, Highest. That's it. amd-pstate, however, while still having the three state limitation, is able to change the frequency of those available states.

This is literally from my Zen 2 PC right now:

sudo cpupower frequency-info

analyzing CPU 3:
driver: amd-pstate-epp
CPUs which run at the same hardware frequency: 3
CPUs which need to have their frequency coordinated by software: 3
maximum transition latency: Cannot determine or is not supported.
hardware limits: 400 MHz - 4.31 GHz
available cpufreq governors: performance powersave
current policy: frequency should be within 400 MHz and 4.31 GHz.
The governor "performance" may decide which speed to use
within this range.
current CPU frequency: Unable to call hardware
current CPU frequency: 3.07 GHz (asserted by call to kernel)
boost state support:
Supported: yes
Active: yes
Boost States: 0
Total States: 3
Pstate-P0: 3700MHz
Pstate-P1: 1700MHz
Pstate-P2: 1400MHz

It supports anything between 400MHz and 4.31GHz. Using acpi_cpufreq P2 would be forced to 400MHz and P0 would be forced to 4.31GHz. I honestly don't remember what my CPU's Middle frequency is to tell you what P1 would be. amd-pstate can dynamically change those three available Pstate frequencies to give the scheduler more room to work with. If you're not running a maxed out workload you can be better served somewhere between Middle and High frequency.

[edwaleni]

I don't have any of the relevant chips, but are you sure you're using amd-pstate? As I recall a very small number of states is what you see with the acpi-cpufreq driver.

Yes, I am running the AMD p-state driver.

[edwaleni]

3 states is an ACPI Ryzen thing. AFAIK, all Ryzens are like this (at least Zen 2 and up on AM4):

With acpi_cpufreq you'll get 3 states. Lowest, Middle, Highest. That's it. amd-pstate, however, while still having the three state limitation, is able to change the frequency of those available states.

This is literally from my Zen 2 PC right now:

sudo cpupower frequency-info

analyzing CPU 3:
driver: amd-pstate-epp
CPUs which run at the same hardware frequency: 3
CPUs which need to have their frequency coordinated by software: 3
maximum transition latency: Cannot determine or is not supported.
hardware limits: 400 MHz - 4.31 GHz
available cpufreq governors: performance powersave
current policy: frequency should be within 400 MHz and 4.31 GHz.
The governor "performance" may decide which speed to use
within this range.
current CPU frequency: Unable to call hardware
current CPU frequency: 3.07 GHz (asserted by call to kernel)
boost state support:
Supported: yes
Active: yes
Boost States: 0
Total States: 3
Pstate-P0: 3700MHz
Pstate-P1: 1700MHz
Pstate-P2: 1400MHz

It supports anything between 400MHz and 4.31GHz. Using acpi_cpufreq P2 would be forced to 400MHz and P0 would be forced to 4.31GHz. I honestly don't remember what my CPU's Middle frequency is to tell you what P1 would be. amd-pstate can dynamically change those three available Pstate frequencies to give the scheduler more room to work with. If you're not running a maxed out workload you can be better served somewhere between Middle and High frequency.

So I was still correct. 3 p-states, you just exposed a freq variation available in each.

There is a table of all the states available in the p-state driver if the Zen level supports it. (sorry, I don't have the link right in front of me)