Exploring Spiking Neural Networks for Prediction of Traffic Congestion in Networks-on-Chip

Networks-on-Chip (NoC) is the most modular and scalable solution for next generation hardware communication where significant data traffic loads are shared across many communication paths. One key challenge in maximising NoC performance is traffic congestion. The management of congestion at the earliest stage can significantly minimize the impact on NoC throughput. Prediction of NoC congestion offers a pre-emptive strategy in maximising NoC throughput. This paper proposes a novel spiking neural network (SNN) approach to prediction of traffic congestion. The proposed SNN exploits the temporal nature of the traffic to identify congestion patterns. The proposed SNN explores two models and both are trained and evaluated to predict local congestion 30 clock cycles in advance of occurring. Results shows that the SNN predictor utilizes 9 times less hardware area than previous approaches and can achieved up to 96.59% in accuracy.