H2H: Heterogeneous Model to Heterogeneous System Mapping with Computation and Communication Awareness
TimeWednesday, July 13th2:15pm - 2:37pm PDT
Location3005, Level 3
AI/ML Design: System and Platform
DescriptionHeterogeneity becomes a trend in both machine learning (ML) models and hardware systems. For algorithm, the heterogeneity in ML models comes from multi-sensor perceiving and multi-task learning (multi-modality multi-task (MMMT) models), resulting in diverse DNN layers and computation patterns. For system, it becomes prevailing to integrate dedicated acceleration components into one system. It introduces a new problem, heterogeneous model to heterogeneous system mapping, which must consider both computation and communication efficiency. We propose a novel mapping algorithm with computation and communication awareness. By slightly sacrificing computation efficiency, the communication latency is largely reduced and the system overall performance is improved.