pgerassi (pgerassi2012.delete@this.wi.rr.com) on August 13, 2018 7:34 pm wrote:
> anon (spam.delete.delete@this.this.spam.com) on August 13, 2018 3:25 pm wrote:
> > anon (delete.delete@this.thank.you) on August 13, 2018 2:48 pm wrote:
> > > Per Hesselgren (perhesselgren.delete@this.yahoo.se) on August 13, 2018 8:50 am wrote:
> > > > Threadripper on Anandtech
> > >
> > > Looking at these graphs from Anandtech's review:
> > >
> > > https://images.anandtech.com/doci/13124/Uncore%207980XE.png
> > > https://images.anandtech.com/doci/13124/IF%20Power%202990WX.png
> > > https://images.anandtech.com/doci/13124/IF%20Power%20EPYC.png
> > >
> > > Intel's Mesh seems to be much more energy-efficient than AMD's Infinity Fabric.
> > > But how did they measure the "core vs uncore" power consumption of AMD and Intel CPUs?
> > >
> >
> > I wouldn't compare the 7980XE and 2990WX but rather the 2950X. The power consumption scales with the
> > core count so only 18 vs 16 cores makes sense, both 18 vs 32 and 18 vs 8 are equally pointless.
> > https://images.anandtech.com/doci/13124/Uncore%207980XE.png
> > https://images.anandtech.com/doci/13124/IF%20Power%202950X.png
> >
> > The real question is does the IF figure include L2/L3 intra CCX traffic?
> > Because if that still has to be added on top of it all the IF is seriously
> > power hungry will the cores themselves are impressively efficient.
>
> It looks like they took values from CPU registers for both core power and uncore power. This without
> verifying that the numbers they got matched reality. So these numbers could be off the real numbers on
> all these CPUs and in different directions. Intel could be underestimating them and AMD overestimating
> them. It stinks of bias without a reality check. Sum the Vcore times Icore per pin and the same for
> all power pins to all other subsystems. Then check each against what the registers say it is.
>
> Trouble is that L3 power comes not from the Vcore pins. Nor does the 128 PCIe lanes, the memory controllers
> (attached or not), the SOC stuff, the PLLs, clocks, etc. Threadripper has far more of that than 7980XE.
> There is still 8 memory channels, even if 4 are unattached and 128 PCIe lanes even if 64 are unattached,
> etc. And 64MB of L3 is a lot larger than 24.75MB of the 7980XE. Of course as a future enhancement to
> fuse off all of the unneeded stuff, removing it from using power. But this does leave some power wasted
> even if the numbers are accurate. In previous reality checks, Intel has been caught underestimating
> true power use while AMD over estimates same. This bias, leads to bad conclusions. Like always, empirical
> results trumps theoretical results. Heck the 2990WX registers might have theoretical use from unattached
> stuff even if its fused off and not using any power at all. Anandtech just sums it up without checking
> that those registers should be ignored (holds garbage values).
>
> "Hoist by their own petard!" Assumptions are the mother of all f*****s!
>
> Pete
>

Well the L3 cache and memory channels do matter for performance though so it would be equally unfair to not count their power consumption while counting the performance they add.

The registers not being accurate is a good point though. We have no idea what AMD and Intel respectively consider core and uncore, especially for the caches it's not clear where they belong and the "SoC stuff" may or may not be considered part of the uncore.