Re: on the verge of shopping for new desktop hardware, recommendations?
On Mon, 2021-03-08 at 21:16 +0100, Linux-Fan wrote:
> Tixy writes:
>
> > On Mon, 2021-03-08 at 05:36 -0500, The Wanderer wrote:
> > > On 2021-03-07 at 22:53, Felix Miata wrote:
> > >
> > > > Linux-Fan composed on 2021-03-08 03:35 (UTC+0100):
> > > >
> > > > > Wrt. power I usually start from CPU + GPU
> > > >
> > > > I used an online calculator
> > > > https://www.bequiet.com/en/psucalculator
> > > > on this system with
> > >
> > > Applying the power-consumption handwave estimates from Linux-Fan's
> > > larger mail, that would be:
> > >
> > > > i3-7100T (TDP 35W) supports up to 3 displays up to 4096x2304
> > >
> > > 35
> >
> > TDP is Thermal Design Power, it doesn't mean the max power.
> >
> > Whilst experimenting with my new desktop, with a 65W TDP CPU, the power
> > of the whole system (measured at the mains supply) went from 16W idle
> > to 170W by executing "while : ; do : ; done" for each core. I'd guess
> > the bulk of that is the CPU, not the memory system.
>
> Yes, this is good to keep in mind. For my system, the power consumption went
> from 100W idle to 264W with your test, i.e. +164W which is close to the
> processor TDP of 165W. Using a different "benchmark" [1], I achived 288W
> i.e. more than the TDP. Of course, the UPS' power display is slow and hence
> I would not notice the real spikes :)
>
> CPU Manufacturers do not seem to publish max power figures AFAICT. Hence it
> seems best to estimate the additional power needed based on experience/tests?
>
> > It soon drops to 130W as the thermal protection kicks in, but you'd
> > want a PSU to cope with the peaks. And maxing out all cores isn't just
> > a theoretical exercise, transcoding video files or compiling programs
> > will happily do that.
>
> The interesting thing with modern CPUs is that even applications that
> seemingly cause 100% CPU usage (like my benchmark [1]) do not actually
> stress the CPU most -- I still do not know all the details about that,
> though. In part, it seems to be related to some instructions needing more
> power than others (vector extensions are known to be power-hungry).
I agree, though as my test got CPU temperature from 40C to the critical
limit of 100C in about 5 seconds, I'd suggest that more power hungry
tests instructions wouldn't actually draw more power, because the
temperature can't rise any more. The thermal protection would just
reduce clock frequency, or increase idle time, more aggressively to
keep power dissipation (and hence temperature) at the limit.
Actually, in my case, the firmware didn't seem to be doing a good
enough job because I was doing these tests to prove the suddenly hard
resetting I was getting using the computer was due to overheating.
I ended up installing 'thermald' package, to throttle the system to
prevent overheating, as the firmware didn't seem to be doing a good
enough job.
--
Tixy
Reply to: