[Beowulf] /. Cooler room or cooler servers?

Jim Lux James.P.Lux at jpl.nasa.gov
Thu Apr 7 19:44:25 EDT 2005


At 04:27 PM 4/7/2005, Mark Hahn wrote:
> > > > No mention of PowerPC which runs far cooler (1/2?) than Intel/AMD:
> > >
> > >except that it doesn't.
> > >
> > > > http://www.terrasoftsolutions.com/products/apple/power-useage.pdf
> > >
> > >lovely graph.  alas, the numbers are pure fantasy.  yes, I actually
> > >did measure a dual xserve g5 (not heavily configured) and it was 200-210W,
> > >quite comparable to dual-opterons and even dual-xeons.
> >
> >
> > Leaving aside such tantalizing issues as what is a GHz, in terms of useful
> > work.
>
>is there some reason to believe I'd deliberately arrange an unfair
>comparison?  in any case, these factors don't matter that much.
>the point is that the true dissipation of 200-210W is drastically differt
>from the macophile reputation of these machines.  at the time, what
>I had to compare with were duals like HP Dl145 (8G, 1.8's).  iirc
>those are around 220W; the same machine with a 2.2 peaks at 240.
>obviously, the new opteron rev would be noticably cooler.


This what I would imagine off hand...  Comparable computational power of a 
given IC fabrication technology (feature size/voltage/etc) will dissipate 
similar amounts of power. And this is what you said you measured.



Nobody has any special physics, so differences will come in the suitability 
of processor/computer architecture (data path widths, cache sizes, etc.) to 
the computational problem being attacked.  And, to a lesser degree, to the 
sophistication of the design, in terms of parts selection, etc.

For one thing, most PCs have fairly inefficient power supplies.  Why bother 
eking out a few percent better Joules/calculation, when the power supply is 
only 70% efficient.  You'd be better off spending your pennies on a 80% 
efficient power supply.

And, for what it's worth, lower voltages may make the processor lower power 
(for a given amount of computational work), but you might lose overall, 
because it's less efficient to regulate to low voltages, and the IR losses 
in ancillary components (PC board traces, solder joints) are higher.  This 
might be a worthwhile trade for a manufacturer, since it might make their 
thermal design easier.



> > However, if we do a straight watts/GHz, Xeon is 34.375, G4 is 33.8 ... not
> > too much difference.
>
>I'm not sure where those numbers come from - I hope you're not just using
>Intel's TDP numbers (which are not max-dissipation numbers.)

I used the (fantasy) ones off the graph in the linked pdf.  Divided Watts 
by GHz, on the simplistic assumption that 2GHz does the work exactly twice 
as fast as 1GHz.

The real point is that the numbers for the biggest and smallest bars in the 
chart (I didn't bother to do the rest) are essentially the same, within a 
few tens of percent.



>W/GHz is amusing; here are some other numbers:
>http://images.anandtech.com/graphs/intel%20dual%20core%20preview%20part%202_040605110408/6838.png
>
>                 w       ghz     w/ghz
>amd             143     2.2     65
>intel sc        214     3.0     71.3
>intel dc        242     2*2.8   43.2
>
>of course, the AMD looks even better if you scale by performance rather 
>than clock.
>
> > One also needs to look at important ancillary issues like power 
> consumption
> > of the cache, bus drivers, etc.  And, of course, instruction stream 
> makes a
> > big difference.
>
>kind of obvious, don't you think?  when I said peak, I meant peak,
>and I'm quoting power dissipation measured at the wall plug (kill-a-watt).
>for these systems, running two compute-intensive programs maxes out;
>messing with memory or disks only adds a few watts (percent).

James Lux, P.E.
Spacecraft Radio Frequency Subsystems Group
Flight Communications Systems Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
tel: (818)354-2075
fax: (818)393-6875


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list