Presentation
MARS: Exploiting Multi-Level Parallelism for DNN Workloads on Adaptive Multi-Accelerator Systems
DescriptionAlong with the fast evolution of DNNs, the hardware system is also progressing rapidly. As a promising solution achieving high scalability and low manufacturing cost, multi-accelerator systems widely exist in data centers, cloud platforms, and SoCs. Thus, a challenging problem arises: selecting a proper combination of accelerators from available designs and looking for effective mapping strategies. To this end, we propose a novel mapping framework, MARS, that can perform computation-aware accelerator selection, and apply communication-aware sharding strategies to maximize parallelism. Experimental results show that our algorithm can achieve great latency reduction on typical workloads, compared to the baseline mapping algorithm.
Event Type
Research Manuscript
TimeTuesday, July 11th4:40pm - 4:55pm PDT
Location3006, 3rd Floor
Design
AI/ML System and Platform Design