Hi
Vaibhav
The simplest answer is that you should try it. Just send 4 replicas to
one GPU and see what happens.
My expectation is that in principle it should work, but much slower than
if you had 1 gpu per replica. Still faster than the cpu MPI version.
However, for a system of your size, the questions becomes if your
particular GPU has enough memory to run 4 replicas, each 60K atoms.
The only thing you can do it try. If there is not enough memory in your
GPU, you will know within seconds.
Adrian
On 8/25/25 5:46 AM, Vaibhav Bhatt via AMBER wrote:
> [External Email]
>
> Dear All,
> I am trying to run H-REMD with 4 replicas. I intend to keep the time of
> simulation equal to 100 ns per replica. Running the calculation on CPU with
> sander.MPI or pmemd.MPI is essentially time-consuming with approximately
> 60,000 atoms per replica. What are the implications of running H-REMD on a
> single GPU node?
>
> Regards,
> Vaibhav Bhatt,
> Ph.D Scholar,
> Department of Chemistry,
> Shiv Nadar Institution of Eminence.
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
--
Dr. Adrian E. Roitberg
Frank E. Harris Professor
Department of Chemistry
University of Florida
roitberg.ufl.edu
352-392-6972
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Aug 25 2025 - 09:30:02 PDT
