Ryzen Mobile Power/Performance With Linux 6.3's New AMD P-State EPP Driver

[kn00tcn]

probably a silly question, but do virtualized guests use the host's governor? want to assume yes, any combination of linux/windows/bsd hosts/guests

[Anux]

And it absolutely does: https://bugzilla.kernel.org/show_bug.cgi?id=215729

This case is not about idle but light load and if there is a bug keeping the lowest freq at 3.9 GHz it has an effect ofcourse. But that's a bug, typically you don't see a difference even under light load.

[aviallon]

probably a silly question, but do virtualized guests use the host's governor? want to assume yes, any combination of linux/windows/bsd hosts/guests

Yes, and Michael recently wrote an article about a patch aiming to improve the situation by sharing scheduler data between the host and the guest. Kvm cpufreq iirc.

[agd5f]

Question, since you might actually know -- does the firmware know about the iGPU? Based on the gaming tests on page 3, in mailbox/uncapped rendering, both epp=balance_performance and epp=performance pick max frequency and epp=power chooses some lower frequency. But what will they do with vsync on?

The firmware manages power for all IPs on the APU, both the CPU cores and the iGPU. The CPU and the iGPU clocks are largely managed independently (pstate driver for CPU and GPU driver for iGPU). The firmware takes into account common limitations that are relevant to all IPs (power, thermals, etc.).

[yump]

The firmware manages power for all IPs on the APU, both the CPU cores and the iGPU. The CPU and the iGPU clocks are largely managed independently (pstate driver for CPU and GPU driver for iGPU). The firmware takes into account common limitations that are relevant to all IPs (power, thermals, etc.).

If they're independent, is that a problem if GPU work is not pipelined with CPU work? Consider:

while(true) {
  simulate();  //5 ms of work on CPU at current frequency
  render();    //5 ms of work on GPU at current frequency
}

Both CPU and GPU see 50% utilization. Lowering CPU frequency will make loop take longer and reduce GPU utilization, then GPU governor might see lower utilization and decide to clock down. Or it could go the other way -- raising CPU freq increases GPU workload and prompts it to clock up. The same situation applies vice-versa from the point of view of the GPU governor.

Seems like potentially unstable positive feedback loop?

In this situation, an informed human would want both CPU and GPU to maximize frequency within power/thermal constraints, in whatever ratio gives globally optimal framerate (maybe same voltage?).