I'd suggest at least one core per GPU and maximal bandwidth between all the GPUs. The bottleneck here is not GPU scaling as a single C2050 already attains ~50% of the performance the entire NICS Kraken supercomputer but rather transferring shared data between the GPUs. The faster that goes, the more GPUs you will be able to throw at the problem.
As long as you have a reasonably modern CPU, core i7 versus quad makes little difference in my experience for single GPU. What it does for multi-GPU where the CPU is used to stage copies between the GPUs remains to be seen.
Scott
-----Original Message-----
From: amber-bounces.ambermd.org [mailto:amber-bounces.ambermd.org] On Behalf Of Ross Walker
Sent: Wednesday, May 26, 2010 16:01
To: 'AMBER Mailing List'
Subject: RE: [AMBER] CPU core requirements for GPU Amber?
Hi Jason,
> Will it then be using 4 CPUs, or will it just stick with 1? I knew the
Yes, it will use MPI and thus will have 1 MPI thread for each GPU being
used.
> multi-GPU patch was coming, I guess I just assumed it would be 1 CPU
> directing x GPUs. Will this support GPUs on multiple machines, or just
> on
Technically there will be no difference between GPU's in a single box and
GPUs in multiple boxes from a theoretical standpoint. From a performance
point of view it is too early to tell.
> the same machine? I'd have been inclined to guess the former, though
> I'd
> heard rumor of an nVidia "supercomputer" that boasts several machines
> with
> several GPUs linked together with an interconnect on steroids
> (obviously not
> a detailed rumor :)).
I couldn't possibly comment on this... ;-)
> All the best (to copy your signature sign-off),
Imitation is the highest form of flattery. ;-)
Ciao!
Ross
/\
\/
|\oss Walker
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
-----------------------------------------------------------------------------------
This email message is for the sole use of the intended recipient(s) and may contain
confidential information. Any unauthorized review, use, disclosure or distribution
is prohibited. If you are not the intended recipient, please contact the sender by
reply email and destroy all copies of the original message.
-----------------------------------------------------------------------------------
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed May 26 2010 - 16:30:06 PDT
