Neuromorphic Computing for Traffic Prediction in Networks-on-Chip using Spiking Neural Network

Network-on-chip (NoC) provides a comprehensive data communication solution in multicore system-on-chip. NoC maximizes throughput/latency performance by offering multiple data communication paths between routing nodes. Unfortunately, NoCs are unable to achieve their full performance potential because of on-path local congestion. NoC requires a predictive solution that reacts to the potential hotspots before their occurrence by rerouting the routed data packets through alternative (minimal) paths, thus minimizing the chances of a local hotspot. This work considers two explicit spiking neural network–based congestion prediction models, router- and network-level, to explore an efficient and low-cost solution to local congestion. The proposed congestion prediction models utilize temporal buffer occupancy patterns generated using trace-based synthetic and multimedia applications to predict local congestion 30 clocks in advance. The models utilize input buffer utilization patterns to predict the local congestion of each NoC router and communicate predictive status with the adaptive routing algorithm of neighboring nodes to consider potential hotspot threats in routing decisions. Compared to existing congestion prediction models, the result shows that both have delivered better prediction performance and require a fraction (0.08%–0.645%) of additional Complementary Metal-Oxide-Semiconductor (CMOS) hardware area overhead to improve network latency/throughput up to 9.99%. This work aims to provide a reliable, efficient, and low-cost prediction model to improve average network latency and throughput performance in a mesh NoC.