Re: [AMBER] REMD on one GPU

From: Ross Walker <>
Date: Fri, 24 Jan 2014 13:51:19 -0800

Hi Milo,

There is nothing to prevent you running multiple replicas on 1 GPU except
that you have to make sure the GPU is not in process exclusive mode -
Check the mode that nvidia-smi reports. And if it is not 'Default' then
run nvidia-smi -c 0 as root.

That said the performance is likely to be very bad and all replicas will
be memory resident on the GPU. So if you had 4 replicas it would use 4
times more GPU memory than if you had 1 replica per GPU. So I would say
that for testing purposes it works but for production runs it will be
difficult to do extensive sampling unless you have unique GPUs for each

I hope that helps.

All the best

On 1/24/14, 1:28 PM, "Milo Westler" <> wrote:

>Can REMD be run (with pmemd.cuda) on a single GPU or do you need a
>GPU for each replica?
>-- Milo
>National Magnetic Resonance Facility at Madison
> An NIH-Supported Resource Center
>W. Milo Westler, Ph.D.
>NMRFAM Director
>Senior Scientist
> and
>Adjunct Professor
>Department of Biochemistry
>University of Wisconsin-Madison
>433 Babcock Drive
>Madison, WI USA 53706-1544
>PHONE: (608)-263-9599
>FAX: (608)-263-1722
>AMBER mailing list

AMBER mailing list
Received on Fri Jan 24 2014 - 14:00:02 PST
Custom Search