There's not really any generic "accelerator" infrastructure in Open MPI itself -- there's a bunch of explicit CUDA support.
But even some of that moved downward into both Libfabric and UCX and (at least somewhat) out of OMPI. That being said, we just added the AVX MPI_Op component -- equivalent components could be added for CUDA and/or AMD's GPU (what API does it use -- OpenCL?). That being said, I would imagine that the data inputs would need to be very large to make it worthwhile (wall-clock execution-wise) to offload MPI_Op operations to a discrete GPU on the other side of the PCI bus. On Feb 11, 2021, at 1:02 PM, Heinz, Michael William <michael.william.he...@cornelisnetworks.com<mailto:michael.william.he...@cornelisnetworks.com>> wrote: Pretty much, yeah. From: Jeff Squyres (jsquyres) <jsquy...@cisco.com<mailto:jsquy...@cisco.com>> Sent: Thursday, February 11, 2021 12:58 PM To: Open MPI Developers List <devel@lists.open-mpi.org<mailto:devel@lists.open-mpi.org>> Cc: Heinz, Michael William <michael.william.he...@cornelisnetworks.com<mailto:michael.william.he...@cornelisnetworks.com>> Subject: Re: [OMPI devel] Support for AMD M100? On Feb 11, 2021, at 12:23 PM, Heinz, Michael William via devel <devel@lists.open-mpi.org<mailto:devel@lists.open-mpi.org>> wrote: Has the subject of supporting AMD’s new GPUs come up? We’re discussing supporting it in PSM2 but it occurred to me that that won’t help much if higher-level APIs don’t support it, too… You mean supporting the AMD GPU in the same way that we have CUDA support for NVIDIA GPUs? -- Jeff Squyres jsquy...@cisco.com<mailto:jsquy...@cisco.com> -- Jeff Squyres jsquy...@cisco.com<mailto:jsquy...@cisco.com>
