Hi Frank,
Provisional benchmarks are attached. All calculation happens on GPU. GPUs
need to communicate when running in parallel. Hence they ALL need to be in
x16 slots and you really need QDR IB to scale between nodes. You need 1 core
per GPU thread - Parallel runs use single sided MPI v2. Optimum is 1 GPU x16
and 1 QDR (Mellanox with GPU Direct) IB card per node (also in x16). 2 GPUs
per node can work although not quite as good. Don't have numbers for 4 gpus
in 1 node but hope to get them soon.
Updates to the webpage
http://ambermd.org/gpus/ will be coming real soon.
All the best
Ross
> -----Original Message-----
> From: "Frank X. Vázquez" [mailto:fxv.umich.edu]
> Sent: Monday, November 08, 2010 3:17 PM
> To: amber.ambermd.org
> Subject: [AMBER] Amber 11 single and multiple GPU question
>
> Hello Amber users,
>
> I wanted to to know some details about serial and parallel GPU simulations
> on AMBER 11. Does the entire calculation occur on the GPU or is part of
the
> calculation occurring on the CPUs? Does the parallel GPU simulation
involve
> communication between multiple GPUs, and does this have a major effect
> on how well the simulation runs? I was also wondering whether there were
> any published benchmarks using multiple GPUs.
>
> Thank you for your response,
> Frank X. Vázquez
>
>
>
> _______________________________________________
> 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 Mon Nov 08 2010 - 17:30:03 PST
