Re: [AMBER] MD GPU Scaling Performance

From: David A Case <>
Date: Fri, 8 Jan 2021 09:47:42 -0500

On Thu, Jan 07, 2021, Jatin Kashyap wrote:
>I am conducting GPU scaling studies on a system containing around 50k atoms
>with protein-ligand/water system.

>My admin suggested that from their past experience, the highest performance
>has been achieved by using single GPU node with maximum GPU cores on it
>i.e. 4, on their machine.

It's not clear if you admin is basing comments on specific experience with
Amber's pmemd, or on more general experience. You can look at the benchmarks
at the Amber web site. Running a serial GPU should be fine, and you should
definitely make sure that pmemd.cuda is working for you before trying any
runs with pmemd.cuda.MPI. Generally, scaling to more than 1 GPU can be
depending on your hardware, and generally gives rather modest performance

>It is giving 1.3 ns/day and with that speed it will takes months for me to
>perform 200 ns trajectory. Is it normal behavior or am I doing something
>very wrong i.e. shall go above 4 GPU cores etc…

1.3 ns/day is way wrong for a 50K atom system. Look carefully for clues
in the mdout file for a serial GPU run. But without more information, I
don't think anyone on the list will know what your specific problem is.


AMBER mailing list
Received on Fri Jan 08 2021 - 07:00:02 PST
Custom Search