AMD Ryzen 9 9950X Power/Performance With CPU Frequency Scaling Driver Tunables

[phoronix]

Phoronix: AMD Ryzen 9 9950X Power/Performance With CPU Frequency Scaling Driver Tunables

Continuing on with the AMD Ryzen 9000 series Linux benchmarking, today's testing is looking at the performance and power impact of the AMD Ryzen 9 9950X when adjusting the CPU frequency scaling driver, governor, and Energy Performance Preference (EPP) tunable to help look at the performance and power efficiency characteristics of this current flagship Zen 5 desktop processor.

https://www.phoronix.com/review/ryzen-9950x-amd-pstate

[superm1]

Think you can re-run some of these with PPD 0.22? The core performance boost and lowest non-linear frequency changes will probably make the results from different tests a lot more pronounced.

[Michael]

Think you can re-run some of these with PPD 0.22? The core performance boost and lowest non-linear frequency changes will probably make the results from different tests a lot more pronounced.

Probably once 6.11 stable out or after 6.12 merge window material will take a fresh look at things as well, or after getting Lunar Lake initial tests out of the way.

[emansom]

Missing amd_pstate=guided with performance and schedutil governors, those seem to offer the lowest frame latency on my Zen 4 7600 in Wine-GE

[aerospace]

Probably once 6.11 stable out or after 6.12 merge window material will take a fresh look at things as well, or after getting Lunar Lake initial tests out of the way.

In some tests the total energy per run can vary of abiut 10%, 20% in some cases. From a server efficiency standpoint this could be useful

[schmidtbag]

Good progress with a lot of this. Nice to see some power-saving profiles with minimal performance losses.

In some tests the total energy per run can vary of abiut 10%, 20% in some cases. From a server efficiency standpoint this could be useful

And laptops

[xenospace]

It's been interesting to see the new Zen 5 benchmarks, but I've really been hoping to see these benchmarks with the system stack compiled for the Zen 5 target architecture. Lots of new instructions available for CPUs in the last couple of years.

I got really excited when I saw that there would be cachyos benchmarks coming soon, as they offer x86-64-v3, x86-64-v4, and Zen 4 level packages for many things.

Really curious whether there would be a big boost to performance if the kernel, and the rest of the stack, was compiled for one of these higher instruction set levels, or even Zen 5, or better yet, -march=native -mtune=native . It's so crazy to know that most Linux distros are still shipping with kernels and software limited to -march=generic -mtune=generic, no LTO, no -O3, etc. in 2024, when x86-64-v3 has been available since 2013!

Perhaps some benchmarks could wake people up to the rapid changes of v4, zen 4, and zen 5 that have happened in the last handful of years as well.

Anyway, waiting with bated breath for the CachyOS benchmarks. Clear and Azure Linux already show some of the benefits of modern CPU support.

[lejeczek]

These with "per Watt" benchmarks are the best - thanks !!! - perplexing to see is that AMD's own 'special' driver/shed does not perform superiorly to Linux's - is it that Linux kernel devel is best of the best (probably) |& AMD is still not investing seriously in Linux core/fundamentals.

[agd5f]

These with "per Watt" benchmarks are the best - thanks !!! - perplexing to see is that AMD's own 'special' driver/shed does not perform superiorly to Linux's - is it that Linux kernel devel is best of the best (probably) |& AMD is still not investing seriously in Linux core/fundamentals.

Linux is still picking the clocks in both cases. The drivers just expose the ability to select clocks/performance levels to the OS. The OS (in this case the governor) is responsible for telling the driver what performance level it wants. The big difference between the acpi-pstate driver and the amd-pstate driver is that the former only exposes 3 clock levels, while the latter exposes 255 levels. Also, the point of the cpu-freq drivers is generally to save power in common work loads or allow the workloads that need it to perform as well as they need. If you are just running CPU intensive tasks at and looking to see how will they will perform, you'll just end up with the cores pegged at the highest sustainable clocks. Where this stuff become more useful is when you are using your laptop for productivity apps and want the battery to last as long as possible. You don't need max clocks. Lower clocks can accomplish what you need and still provide a snappy user experience. Another useful case is on servers where you may have a mix of non-CPU intensive and CPU-intensive workloads. For the former, you want lower clocks because that's all you need for the apps to be performant, and that leaves thermal and power headroom for the CPU intensive workloads. In the case of CPU intensive tests, the acpi-pstate driver generally performs best because there are only 3 clock levels to pick from, as such the governor gets to the max clock pretty quickly. In the amd-pstate driver, there are 255 levels, so the governor may take longer to get to the max level. Running CPU intensive benchmarks is probably not the best test of how well the governors are doing their jobs.