simtk.org

1) Is multi gpu support (i.e. using 2 or more gpus within a node) present in Openmm 4.0?
1a) If so can it be used with Gromacs.
2) If not (i.e. not present in 4.0 release) do you plan to include support for it sometime soon?

jebalunode wrote:1) Is multi gpu support (i.e. using 2 or more gpus within a node) present in Openmm 4.0?

Yes, you can do this with the OpenCL platform. See chapter 6 of the manual. Basically, you can set the OpenCLDeviceIndex property either to a single device index, or to a comma separated list of indices. So if you set it to "0,1", the computation will be split between devices 0 and 1.

Be aware that the parallelization doesn't scale very well with number of GPUs. There's a big speedup in going from 1 to 2, a small speedup in going from 2 to 3, and negligible speedup (or sometimes it even gets slower) going from 3 to 4. Also, using multiple GPUs isn't supported with implicit solvent.

1a) If so can it be used with Gromacs.

I'm not sure about that. You'll need to check with the Gromacs developers to see if they provide any way to set this property.

Peter

We've now added some multi-GPU performance number to the benchmarks page. Those should give you some idea of what kind of speedup you can expect.

Peter