Event Description
Most of the performance improvements in the past decade came from GPU compute. Most compute-heavy codes thus have a strong incentive to transition from CPU-only to GPU-accelerated frameworks. Luckily, many parallel programming concepts that apply to CPU-only compute can also be exploited on GPUs; thus, OpenMP offload emerged as one of the easiest paradigms for accelerating existing code bases. Unfortunately, not everything that can be done in CPU-only OpenMP can be done on the GPU, and in this talk, these limitations and practical workarounds will be explored. There are also performance considerations that must be addressed, with the partitioning of memory between CPU and GPU cores being a major factor. These considerations will be examined in both traditional discrete-GPU systems and more advanced architectures, such as the AMD MI300A-based SDSC Cosmos.
About the Instructor
Igor Sfiligoi has a long history of helping various scientific groups optimize and accelerate their codes on various architectures. He has been working with datacenter-grade GPU systems for almost a decade and has experienced first-hand the challenges of GPU-acceleration during that period. He has also seen the rise and fall of various accelerator ecosystems over that time and has a deep appreciation of how dynamic this environment has been. He is also one of the co-PIs of the SDSC Cosmos HPC system.