Presentation
Memory and Computation Coordinated Mapping of DNNs onto Complex Heterogeneous SoC
DescriptionThe demand of deploying heterogeneous models to heterogeneous SoC is increasing with the advancements in both algorithm and hardware. However, existing mapping frameworks suffer from inefficiencies for lack of awareness of the computation and data transfer imbalance problem in heterogeneous SoC. In this paper, we propose COMB, a mapping framework that coordinates the memory and computation overhead of heterogeneous accelerators to achieve latency improvement and energy efficiency with two optimizations: dataflow grouping and datapath selection. In evaluation, COMB achieves significant improvement in both energy efficiency and end-to-end latency compared to state-of-the-art works.
Event Type
Research Manuscript
TimeWednesday, July 12th1:40pm - 1:55pm PDT
Location3010, 3rd Floor
Design
SoC, Heterogeneous, and Reconfigurable Architectures