An Important Note On The Alder Lake Mobile Power/Performance With Linux 5.19

[hamishmb]

I think OpenBSD has probably caused a lot of power wastage over time with those settings.

Good to see that the AMD processors are doing okay efficiency wise still. I'd assumed the 5950x wasn't great because it's a high end part, but I'm glad to be wrong - I'll probably end up using one to replace my 3600 some time in the future.

[jakubo]

Have I missed something? How is finishing the task at 3/4 of the time and then potentially going into sleep less energy efficient? Or are we seeing drawbacks of too aggressive clocking like, idk, flushing caches repeatedly in 5.18?

[hamishmb]

I may have misunderstood the previous message, but if it just runs at 100% clocks all the time when plugged in to mains, even when idle, that will waste a lot of power.

[yump]

Based on the large variation in the graphs, it looks like maybe the 5.18 kernel is thrashing against some kind of limit that doesn't work gracefully when asked to operate continuously. Might even be something outside the CPU like the voltage regulator temperature.

It'd be interesting to see if cranking the power or frequency limits down enough makes that behavior go away. And if it does, what if any performance difference remains between the kernel versions with both tuned to the same power draw.

True, even on my decade-old Intel i5-3210M (Ivy Bridge) disabling the meager 600 Mhz turbo-boost (2.5 to 3.1 Ghz) would make sure that the fan on my notebook would never reach its highest & loudest spin-state, even under full load with the performance governor.

With that said, disabling the turbo-boost on my desktop-class Intel i7-11700F would cut the maximum frequency from upto 4.9 Ghz down to just 2.5 Ghz, which is just too big of a hit to performance.
Thankfully, cooling these parts inside a PC tower is alot easier than inside thin 'ultrabooks'.

Luckily there's a lot more granularity available than turbo on vs turbo off. You should be able to give your 11700F a 200 MHz haircut with:

cpupower frequency-set -u 4.7GHz

[yump]

Have I missed something? How is finishing the task at 3/4 of the time and then potentially going into sleep less energy efficient? Or are we seeing drawbacks of too aggressive clocking like, idk, flushing caches repeatedly in 5.18?

Because dynamic power scales like frequency cubed or a little worse, which makes energy about frequency squared. Race-to-sleep is overhyped, unless there are large static loads that you can turn off when you reach sleep, like the wifi radio or display backlight.

AMD's term for the highest clock frequency where race-to-sleep actually works, considering only the CPU, is "Lowest non-linear frequency". On some machine that Ray had around the lab, that was 1.1 GHz, for a chip capable of boosting to 4.68 GHz.

[Cryio]

I have a Surface Pro 7 with an i5 1034G4 Ice Lake U, 15W chip. On Windows 10/11, my chip would get stuck to 1.1-1.2 GHz across the vast majority of sustained loads. The Turbo to 3.7 GHz was completely pointless.

On Linux, by default, my CPU could theoretically run at 3.7 indefinitely, given I wasn't power limited, but I lack cooling, so I would get throttled hard to 200 MHz for long periods of time.

I now use a Gnome extension that allows me to limit my CPU clocks. For my device, it seems 2.75 GHz is perfect for any kind of loads given I'm never throttling now and the device is over TWICE faster on Linux than on Windows.

[Eumaios]

For those asking whether the performance will be user adjustable, Michael reports that Intel Idle driver defaults to C1E states for Alder Lake but can be changed to C1 states by entering the parameter "intel_idle.prefer_cstates=2".

[cyring]

C1E perhaps, but there are more to tweak in kernel one can alter with https://github.com/cyring/CoreFreq

See our ADL review at https://gist.github.com/cyring/ef3a3...51b61a4290ee94

[Linuxxx]

Because dynamic power scales like frequency cubed or a little worse, which makes energy about frequency squared. Race-to-sleep is overhyped, unless there are large static loads that you can turn off when you reach sleep, like the wifi radio or display backlight.

AMD's term for the highest clock frequency where race-to-sleep actually works, considering only the CPU, is "Lowest non-linear frequency". On some machine that Ray had around the lab, that was 1.1 GHz, for a chip capable of boosting to 4.68 GHz.

Again, that's all nice in theory, but when Michael's benchmark proofs that Clear Linux with amd-pstate + performance beats all other options by having both the lowest temperatures & lowest power-draw on average, something doesn't seem to properly add up with your expectations:

https://www.phoronix.com/scan.php?page=article&item=hp-devone-linux&num=8

And as I had already told before, I am able to watch a 1440p60 AV1 video on my i7-11700F with intel_cpufreq + performance (Intel energy to performance bias set to 7 in a range of 0-15) without the fan spinning up from its low default state, so clearly race-to-sleep seems to be working as expected here.

[jakubo]

Because dynamic power scales like frequency cubed or a little worse, which makes energy about frequency squared. Race-to-sleep is overhyped, unless there are large static loads that you can turn off when you reach sleep, like the wifi radio or display backlight.

AMD's term for the highest clock frequency where race-to-sleep actually works, considering only the CPU, is "Lowest non-linear frequency". On some machine that Ray had around the lab, that was 1.1 GHz, for a chip capable of boosting to 4.68 GHz.

But we don't have to guess or calculate the power or the time. We measure both directly. All we have to do is sum up all the power values to get the energy per task. its not that hard. If you have a bright back-light, or if you are running a headless server doesn't really matter. Take a job that runs for 5 days and after the job you can turn it completely off and unplug it.
Also... the more power you need to keep the system up, the less impact frequency scaling would have.

Running in high frequencies also would mean that you are not data bound and the CPU runs few idle cycles. So the memory system delivers data close to optimally.

So i have my doubts.