Re: [AMBER] CUDA MPI issues

From: David A Case <>
Date: Fri, 27 May 2016 07:44:24 -0400

On Thu, May 26, 2016, Biplab Ghosh wrote:
> I think, it was a big mistake to buy this system.

Just a general note: It is still the case the parallel scaling in Amber
across many GPUs leads to only modest speedups, even with the best
configurations. I personally just run MD jobs on a single GPU, and take
advantage of the chance to run more individual simulations that way.

Of course, this is not right for everyone. Just be sure that you really need
to run a single job across multiple GPUs. No matter what hardware you get
now, it will become obsolete soon enough. You can take advantage of what you
have learned here when you upgrade to Pascal (or whatever NVIDIA will have in
the future.)


AMBER mailing list
Received on Fri May 27 2016 - 05:00:02 PDT
Custom Search