A Flexible Routing Strategy Based on Region Manager in Network-on-Chip for Neural Network Algorithms

In contemporary neuromorphic acceleration systems, the diversity of neural network models and their corre-sponding network mappings engenders a variety of data flow patterns and transaction sizes. The heterogeneity of these models necessitates tailored routing methodologies. In this paper, we commence with an examination of prevalent traffic patterns in-herent in cutting-edge neuromorphic accelerators. Subsequently, we introduce a routing strategy predicated on configurable region managers. These managers not only ensure deadlock-free routing paths by monitoring network conditions but are also capable of adapting to disparate traffic patterns through the deployment of versatile routing configurations. Through simulations, we have ascertained that our proposed architecture can markedly diminish data latency by 4% to 39% and improve throughput up to 44%, representing a significant enhancement over the conventional mesh network-on-chip (NoC) designs that are currently in widespread use.