Reliable congestion-aware path prediction mechanism in 2D NoCs based on EFuNN

Abstract

The efficiency of networks-on-chip (NoC) is affected by related routing algorithms. This paper aims to develop a reliable routing mechanism in 2D mesh-based NoCs based on evolving fuzzy neural network (EFuNN). Inspiring the advantages of the neural network and the fuzzy system, EFuNN is used for training the network. Online training of EFuNN is faster than that of conventional neural networks trained by the backpropagation algorithm. The proposed algorithm examines distance to the destination, network traffic status, and link faults of the area around the current switch obtaining the reliability metric; it chooses the best path from available paths using fuzzy neural network. Training the network to respond to the routing requests at a specific time is an advantage of using EFuNN. Experimental results show that the on-chip network utilizing the proposed method outperforms state-of-the-art research works in terms of reliability, latency, and throughput regarding different traffic patterns for various 2D mesh topologies.