Intel Core i9 12900K P-State Governor Performance On Linux

[Anux]

I think that shedutil it just needs more Info from the OS and User what is important. You don't need to constantly evaluate, only Look at phoronix every now and than.

[mppix]

Yeah, this has been my suspicion for a while now and therefore it's good to see it confirmed with hard data by Micheal.

But when pondering about it, it does make sense actually:
With the performance governor, your CPU is only faced with a binary choice - either go full speed on a task or idle in the deepest sleep-state possible, thus saving the most amount of energy & heat output.

Dynamic schedulers such as "schedutil" however have a far more complex choice to make:
Trying to second-guess the most 'optimal' frequency for any given task, which in real-world usage scenarios means constantly chasing behind ever-changing workloads while taking longer to complete them and use more power along the way - therefore being more often wrong than right, as seen with Michael's benchmarks.
(i.e. Trying to guess out of the CPU scheduler's a$$.)

Now, these benchmarks are the final nail in the schedutil coffin for me:
Even after all these years, it seems as though it simply can't be improved upon any further; so from here on out, I won't bother anymore evaluating it and just stick to the performance governor at all times.

But there is still one unfortunate idea that keeps nagging me:
The Steam Deck will be using schedutil by default, and I honestly can't see it perform any better on AMD's APU, because the decision-making logic is the very same flawed one we are witnissing here...

Anyhow, hopefully Micheal can be seeded with a Steam Deck by Valve, so we could get a definitive answer via benchmarks!

Schedutil is under active development and is getting access to p-states

Many cups are unsble to stay in the max performance state for long

[darkbasic]

So intel_cpufreq perf seams to be much more efficient than p-state perf while simultaneously beeing on the same performance level making p-state powersave totaly useless. Or am I missreading it?

Exactly, this makes no sense at all. Intel, why?

[Linuxxx]

Exactly, this makes no sense at all. Intel, why?

You do realize that intel_pstate powersave is still a dynamic CPU governor, contrary to acpi_cpufreq powersave, which is the one that always stays at the lowest frequency, right?

Now, when you also realize that both intel_pstate powersave & intel_cpufreq performance are being tested with the "Energy-Performance Profile" (EPP) of balance_performance, then intel_cpufreq performance being the preferable & just overall better option makes quite alot of sense actually:

The performance governor allows the CPU to finish any given task in the most efficient manner possible by simply racing to idle and thus alternating between full-throttle & deep-sleep, whereas intel_pstate powersave tries to be all "smart" about it, similar to schedutil.

The end result speaks for itself...

[darkbasic]

You do realize that intel_pstate powersave is still a dynamic CPU governor, contrary to acpi_cpufreq powersave, which is the one that always stays at the lowest frequency, right?

That was quite clear from the performance figures.

Now, when you also realize that both intel_pstate powersave & intel_cpufreq performance are being tested with the "Energy-Performance Profile" (EPP) of balance_performance, then intel_cpufreq performance being the preferable & just overall better option makes quite alot of sense actually:

The performance governor allows the CPU to finish any given task in the most efficient manner possible by simply racing to idle and thus alternating between full-throttle & deep-sleep, whereas intel_pstate powersave tries to be all "smart" about it, similar to schedutil.

The end result speaks for itself...

It doesn't make that much sense to me: if in a mix of dozens of different workloads pstate ends up consuming more power while also being slower then what's the point? Which workload is supposed to be suitable in order to give some kind of advantage?

[garloff]

It's obvious who the winner is for most scenarios is: intel_cpufreq performance. Just 2% shy of max achievable performance at <85% power consumption.