Linux 6.5 Now Defaults To AMD P-State "Active" EPP For Modern Ryzen Systems

[Anux]

Any thoughts - 5900X or 5800X3D for gaming?

5800X3D is faster in games, more power efficient, cheaper and you can combine it with the cheapest/slowest RAM.

But that CPU does only make sense if you want a 4080 or 7900 GPU, if you just aim for middle class you might get away with a much cheaper 6 core CPU.

[skeevy420]

5800X3D is faster in games, more power efficient, cheaper and you can combine it with the cheapest/slowest RAM.

But that CPU does only make sense if you want a 4080 or 7900 GPU, if you just aim for middle class you might get away with a much cheaper 6 core CPU.

Yep. IMHO, after the 5800x3D, the best CPU to get for gaming would be the either the 5700x since you'll get the same performance of the 5800x without the need of a more expensive, premium, heat sync or the 5600x if the budget is very tight. I'd never consider the 5800x.

The 5800x3D also makes sense for strategy endgames that make all systems crawl.

[kiffmet]

5800X3D all the way for gaming. Having a full 8-core CCX does also give the normal 5800X an edge over the 5900X in many games.

[murlakatamenka]

But when will the MGLRU be also enabled by default?

Guess it means isn't default but just in-kernel. Hmm

[mathieu.cossette]

kernel features and patches, all that by default will take Linux to the next level for gaming ! keep up the good news Phoronix!

[kiffmet]

Guess it means isn't default but just in-kernel. Hmm

If it's built into the kernel, you can enable it by creating the file /etc/tmpfiles.d/mglru.conf (this may not work on non-systemd distros).

w /sys/kernel/mm/lru_gen/enabled - - - - y
#optional; uncomment line below to change behaviour
#w /sys/kernel/mm/lru_gen/enabled - - - - 1

7 means all features enabled and is the default - valid options are 1, 2, 4 or a sum of any of them. Their function is described in the kernel's documentation. To make sure the settings get applied, regenerate your initrd. The changes take effect after rebooting.

[geearf]

I can't get amd_pstate=active to work on 6.4.1 (I didn't try with previous). I get a "failed to register return -19" and the whole system is really slow. It works with amd_pstate=passive though, but I am not sure if that is the right choice for my 3800X. Any advice appreciated, thanks!

[pb72]

The following is happening:
- https://github.com/torvalds/linux/bl...pstate.c#L1495
- https://github.com/torvalds/linux/bl...pufreq.c#L2871 (-ENODEV = -19)

I would check the output of the amd-pstate-ut [1] test. Alternatively, in case the results from amd-pstate-ut are fine and if the 3800X is indeed officially supported, then you'll want to recompile the kernel with CONFIG_DYNAMIC_DEBUG=y and get a better understanding whats going on.

BTW. On 6.4 I would run with amd_pstate=guided (supported since 6.4), instead of amd_pstate=active.

[1] https://docs.kernel.org/admin-guide/...for-amd-pstate

[espi]

I've been using guided mode + schedutil on a 5900X for months now with great results and good minimum frametimes in games.

If I understood the documentation correctly, you don't need a smart governor if using amd-pstate-epp (amd_pstate on active or guided mode). The driver ignores the input from the governor and just lets the CPU firmware choose the frequency.

You should see the same performance using powersave governor.

Now, I don't really understand what guided mode is supposed to do. According to the documentation is the same as active but without the desired performance hint, which I personally find redundant.

[kiffmet]

espi Governor choice still affects core parking. I'm getting ~7W lower idle power draw with powersave (aggressive core parking), while performance (no parking at all) increases it compared to schedutil. "Zenmonitor" shows this as the non-effective CPU core frequencies dropping down to 550Mhz for longer or being pinned at max. value respectively.

The desired performance hint that comes with 'active' mode is mostly useful for mobile platforms IMO, since userspace can tie that preference to being on battery vs. connected to the wall plug, or even make it dependent on the percentage.

While the firmware makes the performance decision on its own in 'guided' mode, the amd_pstate driver definitely still receives information from the governor. The kernel documentation states that the driver was fine-tuned to work best with schedutil and ondemand.

From what I've seen in the testing data attached to the pull request on the LKML, 'guided' is more efficient and came with the largest benefits and almost no regressions compared to acpi-cpufreq, while also beating 'active' mode, which actually regressed in a few tests compared to acpi-cpufreq. 'guided' is best when it comes to absolute throughput, aswell as performance per watt. That data was from before EPP was a thing though.