[me]
> > On my desktop system, using 0 ("auto") mode works quite well: the
> > system idles around 80 to 95 watts, and when building ports or
> > /usr/src with all 32 CPUs going, gets up to the 320-watt range, all
> > without any OS work required.

On Mon, May 25, 2026 at 8:34 AM Olivier Certner <[email protected]> wrote:
> What's the EPP value you are using?  The default of 0?

Yes, I haven't fiddled with any settings except desired_performance,
at this point.

> Now, the question is: Do you get the same performance with these
> more power-flexible settings compared to the driver's defaults ...

Hard to be sure. My wattage measurement is from a cheap power
meter that the machine is plugged into, and I'm not sure what to use
to measure actual performance on real workloads (I don't have a
nice reproducible setup for proper comparison).

To continue though:

> (which are desired performance set to maximum performance
> (which reported value do you see for that?)

There are 32 separate triples.  For example:

    dev.hwpstate_amd.0.desired_performance: 0
    dev.hwpstate_amd.0.maximum_performance: 206
    dev.hwpstate_amd.0.minimum_performance: 12

with the numbers going up to 31 (`sysctl -a` output
counts down from 31 to 0, which feels backwards, but
obviously it doesn't really matter). But each triple is
partly independent of the others, e.g.:

   dev.hwpstate_amd.3.desired_performance: 0
   dev.hwpstate_amd.3.maximum_performance: 211
   dev.hwpstate_amd.3.minimum_performance: 12

Oddly, writing to dev.hwpstate_amd.0.desired_performance
sets all 32 desired_performance values (to whatever value
I choose, even if it's above the reported maximum). I wonder
whether this would be the case on dual CPU-package systems
like a big EPYC setup.

The maximum "maximum_performance" value is 236 and
the minimum is 166 (device numbers 16 and 17), with most
of them clustered around 200-ish.

All of the "minimum_performance" values are 12.

> I suggest you forget about actual frequencies ...

Well, also they're mostly reported as high 5400-ish
when sampled, even with desired performance set to
auto. (Occasionally one will report closer to 5000.)
So they don't seem to be useful in practice.

Chris

Reply via email to