News: 0001469925

  ARM Give a man a fire and he's warm for a day, but set fire to him and he's warm for the rest of his life (Terry Pratchett, Jingo)

One-Line Patch For Intel Meteor Lake Yields Up To 72% Better Performance, +7% Geo Mean

([Software] 61 Minutes Ago 1 Comment)


[1]

Covered last week on Phoronix was a new patch from Intel that with tuning to the P-State CPU frequency scaling driver was showing big [2]wins for Intel Core Ultra "Meteor Lake" performance and power efficiency . I was curious with the Intel claims posted for a couple benchmarks and thus over the weekend set out to run many Intel Meteor Lake benchmarks on this one-line kernel patch... The results are great for boosting the Linux performance of Intel Core ultra laptops with as much as 72% better performance.

[3]

The patch posted last week for the Linux kernel tunes the Intel P-State Energy Performance Preference (EPP) value for Meteor Lake in the default "balance_performance" state. Adjusting the EPP value from 115 to 64 was enough to show a 19% improvement in performance for some workloads and up to an 11% performance per Watt improvement, as shown by the patch series. A similar patch was also posted last week for using those same values with upcoming Intel Arrow Lake processors. See that article for more background information on this tuning.

[4]

As just a handful of benchmarks were shown with that patch and being curious about the Meteor Lake gains, I ran my own set of benchmarks. Using the Acer Swift Go 14 with Intel Core Ultra 7 155H, I ran a set of benchmarks using the current Linux 6.10 Git kernel and then repeated it with the lone Intel P-State MTL tuning patch applied... One line of code changed and then repeating the same set of benchmarks in their default out-of-the-box state on Linux. This is the only Meteor Lake laptop/system I have after buying the Acer Swift Go 14 back on launch day for Linux testing.

This was otherwise a clean Ubuntu 24.04 LTS install besides swapping out the kernel and sticking to all of the other defaults. Besides looking at the raw benchmark results, the CPU package power consumption was also monitored and thus the performance-per-Watt too. Here are the numbers I am seeing for this one-line kernel patch for Intel Core Ultra 7 "Meteor Lake" performance.



[1] https://www.phoronix.com/image-viewer.php?id=intel-meteorlake-epp-perf&image=meteorlake_patch_1_lrg

[2] https://www.phoronix.com/news/Intel-MTL-EPP-Tuning-64

[3] https://www.phoronix.com/image-viewer.php?id=intel-meteorlake-epp-perf&image=meteorlake_patch_3_lrg

[4] https://www.phoronix.com/image-viewer.php?id=intel-meteorlake-epp-perf&image=meteorlake_patch_2_lrg



I went on to test the program in every way I could devise. I strained
it to expose its weaknesses. I ran it for high-mass stars and low-mass
stars, for stars born exceedingly hot and those born relatively cold.
I ran it assuming the superfluid currents beneath the crust to be
absent -- not because I wanted to know the answer, but because I had
developed an intuitive feel for the answer in this particular case.
Finally I got a run in which the computer showed the pulsar's
temperature to be less than absolute zero. I had found an error. I
chased down the error and fixed it. Now I had improved the program to
the point where it would not run at all.
-- George Greenstein, "Frozen Star: Of Pulsars, Black
Holes and the Fate of Stars"