Following Retbleed, The Combined CPU Security Mitigation Impact For AMD Zen 2 / Ryzen 9 3950X

[coder]

"Power efficiency" is a very misleading metric for desktop PCs. You need to consider idle power usage as well. Most of the time the PC isn't doing anything.

Yeah, but even that's too simplistic. I look at both. For typical usage, it's generally running closer to idle (but never actually idle, if you have some web browser windows open!).

However, when you do a build, a render, play a CPU-intensive game, etc. then power-efficiency tells you (relatively) how much heat is going to be generated, in the process. And stats about power consumption under load is going to tell you about what kind of PSU, cooling solution, case, etc. you need, and that will give you an idea about things like noise, size, cost, etc.

[coder]

idle power usage on his AMD Ryzen 5800X with different governors, and observed an increase of only ~1 Watt with the performance governor while being idle,

Idle is idle. The frequency-scaling governors differ principally in how they behave when it's not idle. When it is idle, they will all tend to run the cores at the minimum frequency, because anything else would be needlessly wasteful.

[geearf]

a few games graphs would be good. or there is no real performance impact for games in general?

I agree, I was looking for those as well. Would this not impact RPCS3 for instance?

[Linuxxx]

I agree, I was looking for those as well. Would this not impact RPCS3 for instance?

Definitely, though I'm not aware if RPCS3 can be automatically benchmarked; are you?

[Linuxxx]

Idle is idle. The frequency-scaling governors differ principally in how they behave when it's not idle. When it is idle, they will all tend to run the cores at the minimum frequency, because anything else would be needlessly wasteful.

I think your statement is not quite correct;
here's my understanding:

When the cores are idling, they are doing so in deep sleep states, meaning the frequency actually doesn't matter in that case, because modern processors are physically powering cores off.

Once there is any work to do, the performance governor will request the highest clockspeed available from the CPU, finish the task at hand in the shortest amount possible for the given core and then return to the deepest sleep-state.

So basically, alternating between 0 & 100% clockspeeds as required.

At least that's how I understood "race-to-idle" ...

[geearf]

Definitely, though I'm not aware if RPCS3 can be automatically benchmarked; are you?

I am not aware of anything on that, Dolphin can but it's not as CPU intensive.

[Anux]

Once there is any work to do, the performance governor will request the highest clockspeed available from the CPU, finish the task at hand in the shortest amount possible for the given core and then return to the deepest sleep-state.

So basically, alternating between 0 & 100% clockspeeds as required.

Coder is right in what he describes, yes the cores go to sleep but not 100% of the idle phase because there are still some light background tasks that wake them up and those get served with the lowest possible frequency.

You can see this behavior with the freqency statistic in powertop being 99% idle and 1% lowest freq +/-