PN-TMS: Pruned Node-fusion Tree-based Multicast Scheme for Efficient Neuromorphic Systems

Ziyang Shen, Chaoming Fang, Fengshi Tian, Jie Yang, Mohamad Sawan

Abstract: A growing demand for low-power and real-time computation is motivating the development of dedicated neuromorphic processors. To maximize scalability and power efficiency, multicore architecture has been broadly applied in existing neuromorphic processors. Nevertheless, mapping a Spiking Neural Network (SNN) on a multicore architecture requires a lot of multicast operations. Conventional routing algorithms like path-based routing and dimension order routing (DOR) lead to a severe overhead in both latency and power. To address these limitations, we propose a novel routing algorithm named Pruned Node-fusion Tree-based Multicast Scheme (PN-TMS). PN-TMS leverages multiple algorithms for route planning, optimizing latency and power simultaneously. Experiment results show that PN-TMS outperforms existing network processors’ routing schemes in terms of both energy consumption and latency, achieves an average energy delay product (EDP) reduction of 38.9%.