On Mon, Jun 05, 2017, Vidhya Srivatsan wrote:
>
> However, this is to give a general perspective of MPI (CPU) and CUDA (GPU):
>
> 1. A GPU version is also a parallel version but will run on the device ;
> one can call for more than 1 device and program can be sophisticated to
> parallelize between GPU's . That said, a job on a single GPU is also a
> parallel process (blocks and grids)
> 2. MPI is a CPU version
> 3. MPI / CUDA is soime sort of heterogeneous computing where some processes
> are done in CPU and some in GPU
>
> Again, this is a general overview of CPU, GPU jobs and my reply is not
> specific to AMBER
As you note, the above summary is not accurate for Amber. Perhaps we are
using terminology incorrectly, but item #2 is the one where our usage diverges
from yours. The pmemd.cuda runs calculations on a single GPU (with
administration from the CPU); the pmemd.cuda.MPI code runs calculations on
multiple GPUs. (Hence item #3 above does not apply to pmemd.cuda.MPI.)
>
> On Mon, Jun 5, 2017 at 12:47 AM, Robert Wohlhueter 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?
The above description is accurate as regards pmemd.
...hope this helps.....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon Jun 05 2017 - 05:30:03 PDT
