Optimized Mapping Spiking Neural Networks onto Network-on-Chip

摘要

Mapping spiking neural networks (SNNs) onto network-on-chips (NoCs) is pivotal to fully utilize the hardware resources of dedicated multi-core processors (CMPs) for SNNs’ simulation. This paper presents such a mapping framework from the aspect of architecture evaluation. Under this framework, we present two strategies accordingly: The first tends to put highly communicating tasks together. The second is opposite, which aims at SNN features to achieve a balanced distribution of neurons according to their active degrees; for communication-intensive and unbalanced SNNs, this one can alleviate NoC congestion and improve the simulation speed more. This framework also contains a customized NoC simulator to evaluate mapping strategies. Results show that our strategies can achieve a higher simulation speed (up to 1.37 times), and energy consumptions can be reduced or rise very limited.