MAMAP: Congestion Relieved Memetic Algorithm based Mapping Method for Mapping Large-Scale SNNs onto NoC-based Neuromorphic Hardware

Abstract

Spiking neuron networks (SNNs) have been widely adopted to solve complex tasks such as pattern recognition, image classification, and natural language processing. Simulating SNNs on many-core neuromorphic hardware has been proved to be powerful and energy-efficient. Constrained by hardware resources, it is difficult to map a complete SNN on a single neuromorphic core. The SNNs need to be segmented and placed in different neuromorphic cores. The Network-on-Chip (NoC) connection strategy for interconnecting multiple neuromorphic cores has been widely used in modern neuromorphic hardware. However, the enormous inter-core communication from large-scale SNNs will lead to congestion of NoC and dramatically reduces performance. In this paper, we propose a congestion relieved memetic-algorithm-based mapping method (MAMAP) to find an efficient mapping solution that minimizes the latency and power consumption of neuromorphic hardware by maximizing bandwidth utilization. Besides, our proposed memetic algorithm combining the global search capability of the Particle Swarm Optimization (PSO) algorithm and the local search capability of the TABU, can find the best mapping solution fast. Compared with the state-of-art SpiNeMap, the experimental results show that MAMAP can reduce the average latency by 63% and average energy consumption by 69%.