Multi-GPU Programming is the eleventh and twelfth workshop in this series and was presented by Jiri Kraus (NVIDIA) and Daniel Howard (CISL's Consulting Services Group) on June 30th and July 14th. The video recordings for these session are uploaded here: Part 1 and Part 2. Topics covered include:
Part 1 - Presentation Slides by Jiri Kraus, NVIDIA
- Hardware overview and data movement pathways across nodes and intranode
- GPU DirectCompute Models
- CUDA Aware MPI vs NVSHMEM vs NCCL
- UCX and other specific considerations for Casper and Derecho
Part 2 - Jupyter Notebook
- Interoperability of OpenACC with MPI and NCCL GPU communication libraries
- Hands-On Implementation of MiniWeather with CUDA aware MPI and NCCL
Part 1 of this session featured a presentation by Jiri Kraus, a principle developer technology engineer with NVIDIA. Part 2 of the session features an interactive component in the form of a Jupyter Notebook by Daniel Howard.
The presentation slides plus notebook and a html/PDF version of the notebook is available for use and reference in this directory.