2020 OCP Virtual Summit

OAM or ODSA complaint accelerators have much more efficient interconnection between various accelerators. Therefore they should have much larger potential than the traditional architectures when computation tasks are partitioned and run them on different accelerators in parallel. Due to low communication cost, the software optimization strategies and scheduling heuristics also need to be adjusted accordingly. In this talk, we introduce Sinian, an open source heterogeneous framework to enable heterogeneity-aware computation parallelism to maximize the performance for OCP-compliant accelerators. Task partitioning can be based on efficient architecture cost models, or deeply customized by applications.