Predicting Local Congestion at Fine-grain Levels in Networks-on-Chip Using Spiking Neural Networks

Networks-on-Chip (NoC) was introduced to achieve maximum communication performance in on-chip systems. Local congestion caused by the queuing of data at input channel buffers constrains NoC latency and throughput performance. NoCs require a predictive approach to minimize local congestions. In this paper we proposed a novel fine-grain congestion prediction approach based on Spiking Neural Network (SNN), which predicts router congestion with 30 clock cycles look-ahead capability. Two fine-grain prediction approaches including router and network models are proposed. Prediction performances are evaluated and analyzed using both the synthetic and real-time NoC traffic applications. Results show that the network model has a better performance for fine-grain local congestion prediction, and requires 42% less hardware area than the router model.