GPU Series: Multi-GPU Programming, Part 1

The next session in a series of workshops and tutorials on GPU computing – “Multi-GPU Programming, Part 1” – will be on Thursday, June 16, at 10 a.m. MDT and presented by Jiri Kraus, Principal Developer Technology Engineer at NVIDIA.. The series of one-hour sessions is for scientists, software engineers, and students in the Earth system sciences to help them prepare to use the extensive GPU-computing capabilities of the new Derecho system. The training is intended for both new and more experienced GPU programmers and requires familiarity with the UNIX shell; Fortran, C/C++, and/or Python; and typical program compilation workflows.

Topics to be covered in this session include:

• Hardware overview and data movement pathways across nodes and intranode
• GPU DirectCompute Models
• CUDA Aware MPI vs NVSHMEM vs NCCL
• Task placement
• UCX and other specific considerations for Casper and Derecho

As with other sessions in the series, this one will include dedicated Q&A time and some will involve interactive coding with participants. This session will involve some interactive participation so make sure you are able to access the Casper compute system prior to the event. Recent registrants will be provided login details prior to the session.

Please use this form to register if you plan to attend and have not already registered for the series. You only need to register once, and can then use the same credentials to attend subsequent sessions on the Zoom meeting platform. By registering and attending, you agree to abide by the UCAR Code of Conduct. These training sessions will be recorded and published publicly for future reference.