Abstract

By introducing wireless interfaces in conventional wired routers or hubs, wireless network-on-chip (WiNoC) is proposed to relieve congestion pressure from high volume inter-subnet data transmission. Generally, processing elements on chip receive input data and return feedback through network interface, and data transmission function in Network-on-Chip (NoC) is completed by routers. Hubs equipped with wireless interface are fixed to certain wired routers. While wireless channels may not be fully utilized due to unbalanced workload and constant hub-router connection, e.g., certain nodes processing excess inter-subnet data traffic are far away from hubs. In this paper, we proposed a workload-aware WiNoC design with intelligent reconfigurable wireless interface to improve wireless resources utilization and mitigate congestion. Through multidimensional analysis of traffic flow, a 4-layer neural network is trained offline and applied to analyze workload in each tile, and return three most potential tiles for wireless interface reconfiguration to fully utilize wireless channel and lowing latency. We also implement a historical traffic information-based reconfigurable scheme for comparation. Evaluation results show that in an 8 × 8 hybrid mesh topology, the proposed scheme can achieve 10%–16% reduction in network latency and 5%–11% increment in network throughput compared with fixed-link hub-node connection scheme under several mixed traffic patterns.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call