pmemd.cuda is a separate code from pmemd (Fortran). The CUDA code depends
on the Fortran code to do administrative things, but all of the work is
done on the GPU, including force calculations, particle propagation, and
pressure computations. The CPU orders the GPU "compute the forces," then
"move the particles," and perhaps "rescale the box." The GPU then reports
"done." Multiple GPUs can be employed for these tasks, and the GPUs
communicate over the SLI interface in pmemd.cuda.mpi. I believe the MPI
code also uses multiple CPUs, but this is beside the point.
Dave
On Sun, Jun 4, 2017 at 3:17 PM, Robert Wohlhueter <
bobwohlhueter.earthlink.net> wrote:
> Am I right in concluding that the CUDA version (of pmemd) uses a
> cuda-device, but not multiple cores? That the MPI version used multiple
> cores, but not a cuda-device? And that the MPI-CUDA version will work
> only if there are multiple cuda-devices avaiable?
>
> Bob Wohhueter
>
> _______________________________________________
> 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 Sun Jun 04 2017 - 13:30:02 PDT
