[Beowulf] 90nm opteron power usage

Ben Mayer bmayer at gmail.com
Mon Aug 8 13:15:31 PDT 2005

It would depend on which core you are using in DNet. Most of the cores
are hand optimized so that they sit in cache (at least the RC5 core
does). If you get something that pushes the memory as well as the CPU
you should be able to get the machine to draw more power. And of
course if it grinds on the HD it will draw power, but if your code is
doing that you really are not going any where quickly.

The reason I comment on the memory is that when you read out a block
of memory the  chip has to recharge the items it just read, if they
were 1 (or what ever the charged representation is).

So if you had something that would peg the CPU and read memory with
poor cache access, you should be able to get it, reasonably higher.
Maybe try installing Photoshop and have some fun with large Tiffs.

Anyone know how much the memory could contribute to the overall power load?


On 8/8/05, Bill Broadley <bill at cse.ucdavis.edu> wrote:
> I managed to hook up my kill-a-watt to a dual opteron 2.6 GHz, 4GB pc3200,
> and a 73GB scsi disk (a sun v20z).  The highest power consumption I
> could find was was running 1 distributed.net client per cpu.
> The results was 197-198 watts and 200-201 VA.  I'm quite impressed,
> our 2.2GHz version of the same box is in the 270-300 watt range.
> Has anyone found a (free/open) code that manages higher power consumption
> than the distributed.net client?
> Kudos to AMD for such a big improvement, I thought it might be useful
> for those planning cooling or power for any of the newer 90nm AMD chips.
> Anyone have similar numbers for a new dual nacoma/64 bit xeon?
> --
> Bill Broadley
> Computational Science and Engineering
> UC Davis
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

Benjamin Mayer
University of Minnesota
Ph.D. Student, Computer Science
HPC, Data Mining for Medical Informatics and Network Intrusion Detection

More information about the Beowulf mailing list