On Friday, 28 June 2024 18:37:06 CEST Mario Limonciello wrote:
> On 6/28/2024 11:18, Diederik de Haas wrote:
> >> I don't think so.  I've never heard of this actually used in a desktop
> >> board.  It's for mobile designs AFAIK.
> > 
> > I can understand that the initial/original goal/target was mobile.
> > But is there a(ny) technical reason why it couldn't also support 'desktop'
> > systems?
> > IIRC and IIUC it does need Zen 3, which my CPU/SoC does.
> 
> It needs information about the hardware thermal design to change the
> correct coefficients.

Isn't that something that AMD would know?

> For example just changing the sPPT or fPPT without the appropriate
> thermal headroom will cause the SOC to go into thermal throttling
> prematurely and hurt your performance.
> 
> That's why OEMs encode this information in their EC or in the BIOS and
> advertise it in the PMF ACPI tables what they can do.
> 
> > Ok. I might be able to convince Asus to add it, but I can also configure
> > my system to 'modprobe' it on boot up.
> 
> Loading it with modprobe doesn't do anything if there is no ACPI device
> to bind to.  It's just like having it compiled into your kernel and it
> not binding to any device.

Ok, understood. I'm not too positive I could convince Asus to add that to 
their 'BIOS', but I can try.

> Like I said; I've never seen it on a desktop.  Because of what it's
> intended to do in it's current form, I think it's unlikely that it would
> be added there.

Understood. Hopefully this does make AMD aware that non-laptop/mobile
devices could (or should!) be a potential target too.
Even if it won't directly benefit me now, that would still be a win :)

> >> The way that it works is that the OEM will set bits in their BIOS for
> >> that ACPI device indicating which "AMD_PMF" features they support.
> >> That's things like Static Slider, Auto mode, CNQF, Smart PC, slider
> >> notifications.
> > 
> > I have no idea what those things are, but I can research it ...
> > 
> > But I think everyone and the planet would benefit from a (more) energy
> > efficient system, regardless if it (normally) runs on batteries or not.
> > In my case, this system will likely be my NAS, so idle 99(.9)% of the time
> > and (in the future) on 24/7, so I'm extra motivated to make it consume as
> > little energy as possible.
> > 
> 
> Keep in mind PPD does multiple things based on what the hardware
> advertises support for.
> 
> For your system, you are using amd-pstate in active mode.  I can see
> this from your powerprofilesctl output.  This will make sure the SoC is
> properly biased towards efficiency or performance based on your preference.
> 
> It will operate under the limitations of the coefficients programmed by
> the firmware at max load (so you can't change PPT), but when idle it
> shouldn't consume more power than needed.
> 
> There is a kernel patch series coming in kernel 6.11 for "Core
> performance boost" which will let you turn on/off boost above nominal
> frequency.  You can turn it off, and then CPU cores won't ever go above
> nominal frequency.
> 
> I'm proposing that we would leave boost on for PPD balanced and
> performance states.  For PPD power-saver we would turn it off.
> 
> This is merged in the PPD tree and unless anyone complains will be part
> of the next PPD release.
> 
> The other thing that could be really interesting to do for power savings
> purposes is to put certain tasks only the non-preferred cores or tune
> EPP towards efficiency on cores running long running tasks that don't
> need to finish quickly.
> 
> In my opinion this is going to open up a lot of potential for sched-ext.
> and for systemd.  Userspace can read the core rankings and when starting
> up and moving tasks around adjust EPP, boost and which cores tasks run
> on.  It's a lot of complicated work ahead though with a lot of profiling
> needed.

I hadn't used PPD before, so I'll experiment with it.
Looks like I have lots of research to do ;-)

Thanks for your detailed response :-)

Hack (and Save) the Planet!
  Diederik

Attachment: signature.asc
Description: This is a digitally signed message part.

Reply via email to