AMD P-State vs. ACPI CPUFreq Testing With Ryzen Laptops On Linux 5.17

[phoronix]

Phoronix: AMD P-State vs. ACPI CPUFreq Testing With Ryzen Laptops On Linux 5.17

One of the most prominent features of Linux 5.17 for end-users was the introduction of the AMD P-State driver that is designed to deliver better energy efficiency than the generic ACPI CPUFreq frequency scaling driver relied on by AMD Ryzen processors up to this point. For those wondering how the performance and efficiency currently compare for Ryzen laptops, here are some benchmarks recently carried out on Linux 5.17 for both drivers and testing both the Schedutil and Performance governors.

https://www.phoronix.com/vr.php?view=31024

[tildearrow]

Typo:

For this round of testing I was using a Ryzen 5 5500U (Zen 2) and Ryzen 9 5900HX (Zen 3) latops given they are the latest AMD laptops I have

[wooque]

all this effort to create diver that is same or worse than existing one...

[Anux]

Strange how pstate is faster for one CPU but not the other. But at 10 W less it seems to be a good adversary for powersave.

[potens]

I think it would have been interesting to also include desktop CPU in the comparison because, on my 3900x, I see it hovering waaaaaay lower than it was either with acpi-cpufreq. And, I know some people will say that's useless to do "powersave" on desktop, but, from my point of view, there is no point to keep all cores at 1.2Ghz, when in fact the can be at 560Mhz (which is the kind of difference I'm talking about), because a) electricity cost quite a lot where I live, b) more heat -> fan have to turn faster -> more consumption, and even worse, more noise. (and in fact, I'm very jealous that, on windows, most of the time, a lot of cores are just "off", not slow, but off). But to be honest, I don't use too much schedutils, which for me is good at nothing, so I switch between powersave, conservative, and performance

[Anux]

and in fact, I'm very jealous that, on windows, most of the time, a lot of cores are just "off", not slow, but off

Do you mean C-States? Those are also used in Linux, you can check that with powertop.

[potens]

Do you mean C-States? Those are also used in Linux, you can check that with powertop.

I don't know too much how windows consider it "off" (they just doesn't appears, if I'm remembering well), but, the lower I see with powertop is "C2" (more than 90% on powersave, which is fun, since there is 2% on C1, so, a huge part is missing, I guess it's C0), another fun part is whatever schedutils, conservative or powersave, the percentages in C1 and C2 are almost the same, only the frequencies looks just higher.

PS: I'm not too savy on this specific domain, so I may miss understanding about this (and that's why I'm interested in articles that gives that pre-digested)

[FireBurn]

It would have been useful if all the tests had performance per watt results, simply saying pstates + schedutil wasn't as fast doesn't really give the full picture especially when the percentage difference isn't that much but could make a difference in how long the battery lasts

[geearf]

It would have been useful if all the tests had performance per watt results, simply saying pstates + schedutil wasn't as fast doesn't really give the full picture especially when the percentage difference isn't that much but could make a difference in how long the battery lasts

I was looking for the same, since the whole point is to save battery at best performance.

[jochendemuth]

It would have been useful if all the tests had performance per watt results, simply saying pstates + schedutil wasn't as fast doesn't really give the full picture especially when the percentage difference isn't that much but could make a difference in how long the battery lasts

I think when testing this feature set there are really two overarching questions: what does it bring? what does it cost?

The test series was relatively elaborate so, I would have expected to answer these two questions. But we mostly got an answer to the cost question. As a minor gripe one can put into question the mix of tests run on an end-user laptop... are selenium tests really indicative of a typical end user workload?

The key question though is what is the power consumption difference at idle? Arguably, most end user computers spent most of their time idling or close to it.