Meteor Lake EPP Tuning For Greater Performance Or Power Efficiency With Intel Ultra Core 7

[phoronix]

Phoronix: Meteor Lake EPP Tuning For Greater Performance Or Power Efficiency With Intel Ultra Core 7

With all of my initial Intel Core Ultra 7 155H benchmarking since last week from the Windows 11 vs. Linux performance to the integrated Intel Arc Graphics performance it was at the Linux 6.7 kernel defaults / OS defaults. But for those wanting to push the Intel Core Ultra 7 155H further for better performance or conversely wanting to extend the battery life further for devices like the Acer Swift Go 14 laptop, the Intel P-State Energy Performance Preference (EPP) can be adjusted. Here's a look at the sometimes significant difference to power and performance when adjusting the Intel Meteor Lake CPU between the default balanced performance mode and the alternative power saver and performance modes.

https://www.phoronix.com/review/intel-meteorlake-epp

[pWe00Iri3e7Z9lHOX2Qx]

Thanks for these. I'd probably just run on power-saver exclusively when not plugged in.

[coder]

Holy cow!

In this case, what must've happened is the Power Saver plan put the (single) thread on an E-core. That said, now that I'm looking at it more closely, it seems something is broken in that case, because you can distinctly see 3 bumps in the red graph (the first two are partially obscured), but it appears there's some kind of long stall before the 3rd, where the CPU is just idling.

Michael , could annotate the power, frequency, and temperature plots with the benchmark details, so we can be certain which ones go together?

[coder]

Thanks for these. I'd probably just run on power-saver exclusively when not plugged in.

Well, in some of the more heavily-threaded tests, it hardly seems to make a difference:

PowerSaving didn't even save a whole Watt!


QuantLib (Multithreaded) is another such example.

[loganj]

are Pcores ignored in power save mode?

[Mitch]

Well, in some of the more heavily-threaded tests, it hardly seems to make a difference:

PowerSaving didn't even save a whole Watt!


QuantLib (Multithreaded) is another such example.

Ignoring the 2 LP cores, the processor has 6+8 cores, which means 14 total. The Selenium bench looks single-threaded, or close to single-threaded. If that's true, then the power saver seems to cap Watts per core to about 2.8 watts. 2.8 * 14 = 39.2 watts total, which is already in excess of the blender bench watt usage, which seems multi-threaded. So it seems that the power saver mostly gets its savings from capping individual core usage, not all-core usage. All-core usage must already be capped equivalently in all profiles because the individual core usage seems to peak between 12 watts and 22 watts, which would imply an all-P-core usage of 6*12 (72) to 6*22 (132) watts peak usage for JUST P cores, but we're not seeing anything like that wattage in multi-threaded benchmarks.

In other words, power efficiency seems to be about preventing any single core from using too much power because that is what kills our energy and thermal efficiency. Putting 12 watts into an all-core load makes sense for efficiency. Putting 12 watts into a single core load is counterproductive in terms of efficiency.

[linuxgeex]

In other words, power efficiency seems to be about preventing any single core from using too much power because that is what kills our energy and thermal efficiency. Putting 12 watts into an all-core load makes sense for efficiency. Putting 12 watts into a single core load is counterproductive in terms of efficiency.

Yes, that's always been the case, since the first laptops. This is the furthest thing from new information, lol...

If you've followed overclocking at all, you know that to squeeze the top 2% out of a core, you nearly double its power per performance. The top 30% of the frequency of an Intel K-series typically comes with a 300% power per watt penalty. You can take a desktop K-series and make it sip power like an Atom if you clock it down. However the motherboard will be overdesigned for such low power usage, and so the actual system power usage won't scale down well. To get good efficiency the whole system needs to be designed for it.

Intel are marketing something as new, hoping that we've forgotten mobility 101.

[cj.wijtmans]

Well, in some of the more heavily-threaded tests, it hardly seems to make a difference:

PowerSaving didn't even save a whole Watt!


QuantLib (Multithreaded) is another such example.

power saving works on a "do the workload as fast as possible so we can idle" pinciple so yeah it only affects day to day workloads.

[loganj]

Michael how do you calculate performance/watt? do you take in consideration the longest period of time (power save mode) for all power modes (even idle power consumption after tasks) or only the time it takes for tasks to complete?

[pWe00Iri3e7Z9lHOX2Qx]

The processor has 6+8 cores, which means 14 total. The Selenium bench looks single-threaded, or close to single-threaded. If that's true, then the power saver seems to cap

It's actually 6+8+2 in this case, 16 cores total, 22 threads.

P core base / turbo: 1.4 GHz | 4.8 GHz
E core base / turbo: 900 MHz | 3.8 GHz
LP E core base / turbo: 700 MHz | 2.5 GHz

It would be neat if you could disable the P cores through power-profiles-daemon in a battery save mode, and if there was an extreme battery save mode that powers down everything except the LP E cores (e.g. I'm on a long flight to India and I'm just catching up on email and watching movies and a dual core 2.5 GHz peak is fast enough). I assume I could do it with the old loop like for i in {0..13}; do echo 0 > /sys/devices/system/cpu/cpu$i/online;done.