Ryzen Mobile Power/Performance With Linux 6.3's New AMD P-State EPP Driver

[phoronix]

Phoronix: Ryzen Mobile Power/Performance With Linux 6.3's New AMD P-State EPP Driver

With Linux 6.3 there is the new AMD P-State EPP driver code for supporting the ACPI Energy Performance Preference (EPP) to further enhance the power efficiency and performance of modern AMD systems on Linux. Last week I ran some benchmarks of AMD EPYC with the new AMD P-State EPP mode while in today's article is a look at the laptop impact with Ryzen Mobile when comparing ACPI CPUFreq, the existing AMD P-State driver, and the new AMD P-State EPP mode and its multiple different preferences.

https://www.phoronix.com/review/amd-pstate-epp-ryzen-mobile

[X_m7]

I wonder how the EPP modes compare to acpi-cpufreq and the current passive mode of amd-pstate when it comes to power efficiency/battery life for light to medium workloads like watching a video, casual browsing with a few tabs, typing up documents and so on, currently I end up using amd-pstate's passive mode plus the conservative governor on Linux 6.2 on my Ryzen 7 6800H laptop so it doesn't go full throttle for small stuff and end up causing the second fan to spin up unnecessarily like with the default schedutil governor. The responsiveness difference is noticeable, but it's not like it feels sluggish, especially since I'm coming from a Core i7-6700HQ and a SATA SSD anyway, but it sure would be nice to have an option that still responds quickly for interactive workloads but doesn't go for that 4.7GHz boost clock all the time.

[Zeioth]

I've been using it for half a year or so now. It works great.

[aviallon]

The benchmarks are always appreciated, but as X_m7 says, light workloads and idle power consumption are more interesting for a power-optimizer laptop.
The best governor for raw performance would always be "locked boost maximum frequency", so performance isn't really important here.

What matters is Joules needed to achieve a specific task, and average power when doing light multi tasking (typical company laptop usage).

[edwaleni]

This is an excellent set of test results.

I will be duplicating the test setup on a Ryzen 9 6900HX which has been having random issues when a low power step is done and see if the AMD EPP driver resolves it.

I have been waiting to see some test results before I go invest the time to test. Thank you.

[agd5f]

Doing performance benchmarks is probably not the right way to evaluate this. The goal is not to get the best possible performance; if you want that, just force the clocks to the highest level. The goal is to only use enough power to get the job done while providing a responsive system. The reason the old acpi-pstate driver tends to perform well (i.e., absolute performance) is because it only exposes 3 clocks states, so for the most part the governors either pick the min or the max. The AMD pstate driver in passive mode gives you the full frequency range to choose from, so governors can pick from a wider range of frequencies, as such it takes longer to ramp up to the max. The AMD pstate driver is EPP mode is mostly controlled by firmware. The SMU firmware picks the optimal frequency based on internal perf counters that it tracks, the governor really only provides a hint as what sort of performance they want.

[Linuxxx]

Doing performance benchmarks is probably not the right way to evaluate this. The goal is not to get the best possible performance; if you want that, just force the clocks to the highest level. The goal is to only use enough power to get the job done while providing a responsive system. The reason the old acpi-pstate driver tends to perform well (i.e., absolute performance) is because it only exposes 3 clocks states, so for the most part the governors either pick the min or the max. The AMD pstate driver in passive mode gives you the full frequency range to choose from, so governors can pick from a wider range of frequencies, as such it takes longer to ramp up to the max. The AMD pstate driver is EPP mode is mostly controlled by firmware. The SMU firmware picks the optimal frequency based on internal perf counters that it tracks, the governor really only provides a hint as what sort of performance they want.

On Intel CPUs I can hint to them to prefer maximum energy efficiency via the following:

sudo cpupower set --perf-bias 15

Can I do something similar via the in-kernel cpupower tool on AMD, too?

[Eumaios]

What happened to amd_pstate_epp performance balance_performance from the server test? Is that not supported on the laptop processor?

[Eumaios]

light workloads and idle power consumption are more interesting for a power-optimizer laptop. . . .
What matters is Joules needed to achieve a specific task, and average power when doing light multi tasking (typical company laptop usage).

Does anyone know of any such tests? I searched the internet but found none.

Perhaps a benchmark set up as a multi-hour stream of light office instructions, measuring total power used?

[Michael]

What happened to amd_pstate_epp performance balance_performance from the server test? Is that not supported on the laptop processor?

Not supported on this laptop