Predicting Networks-on-Chip traffic congestion with Spiking Neural Networks

Network congestion is one of the critical reasons for degradation of data throughput performance in Networks-on-Chip (NoCs), with delays caused by data-buffer queuing in routers. Local buffer or router congestion impacts on network performance as it gradually spreads to neighbouring routers and beyond. In this paper, we propose a novel approach to NoC traffic prediction using Spiking Neural Networks (SNNs) and focus on predicting local router congestion so as to minimize its impact on the overall NoCs throughput. The key novelty is utilizing SNNs to recognize temporal patterns from NoC router buffers and predicting traffic hotspots. We investigate two neural models, Leaky Integrate and Fire (LIF) and Spike Response Model (SRM) to check performance in terms of prediction coverage. Results on prediction accuracy and precision are reported using a synthetic and real-time multimedia applications with simulation results of the LIF based predictor providing an average accuracy of 88.28%–96.25% and precision of 82.09%–96.73% as compared to 85.25%–95.69% accuracy and 73% and 98.48% precision performance of SRM based model when looking at congestion formations 30 clock cycles in advance of the actual hotspot occurrence.