Currently, there isn't an MPI fabric/software in existence that enables
scaling past 16 GPUs for multi-GPU simulations. Even for GB on large
molecules, which has plenty of work to do, something very bad happens past
8-16 nodes that causes scaling to peter out.
And without hardware to test on (I don't have reliable access to anything),
there's no way to diagnose why...
On Wed, Aug 1, 2012 at 3:40 AM, Catein Catherine <askamber23.hotmail.com>wrote:
>
> Dear Ross and AMBER experts,
> I noted that AMBER12 is compatible with NV C2050 GPU from your website.
> What is the maximum no. of node tested in parallel mode?
> I used to run my jobs in CPUs with 64 nodes for parallel runs(pmemd),
> what's about GPU?
> I heard it is difficult as the jobs will stop in GPU parallel runs. What
> is the bottle neck for this?
> Best regards,
> Catherine
>
>
> _______________________________________________
> 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 Aug 02 2012 - 08:30:04 PDT
