P2P cannot achieve full interconnect across more than 8 GPUs. You can use
8+ GPUs each on their own for throughput, just not P2P, the communications
protocol used to move data around...
On Tue, Jan 27, 2015 at 5:38 AM, David A Case <case.biomaps.rutgers.edu>
wrote:
> On Tue, Jan 27, 2015, Harmeet Kaur wrote:
> >
> > Can anyone suggest what is the maximum no. of GPUs on which a simulation
> > can be run without compromsing the efficiency? Our institute has 6 cores
> > cluster with 2GPU cards each amounting to a total of 12 cards. Can I run
> a
> > job using all 12 GPUs? I have heard that when one exceeds usage to more
> > than 2 GPUs, the efficiency of the run is decreased. Kindly suggest.
> Thanks
> > in advance
>
> You heard correctly. You can check peformance at the benchmark page:
>
> http://ambermd.org/gpus/benchmarks.htm#Benchmarks
>
> Full details of how to use multiple GPUs is given here:
>
> http://ambermd.org/gpus/
>
> Scroll down to "Considerations for Maximizing GPU Performance". Short
> answer
> is that using 1 or 2 GPUs per run is probably the most you should be using
> with the hardware you describe. (But go ahead and experiment!)
>
> ...dac
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jan 29 2015 - 18:30:03 PST
