Broadcast-Based Hybrid Wired-Wireless NoC for Efficient Data Transfer in GPU of CE Systems

Abstract

Conventional network-on-chips (NoCs) architectures have been adopted as the interconnect network backbone to handle the traffic between many cores and a few memory controllers (MCs) in the graphical processing units (GPUs). Due to the many-to-few/few-to-many traffic pattern, the performance of GPU NoCs is constrained by limited terminal bandwidth at MC routers and traffic hotspots generated in the regions close to MCs. In addition, many GPU applications have redundant data sent from an MC to different cores and this leads to inefficient utilization of limited bandwidth at MCs, which are in fact the network bottleneck. We present a hybrid wired-wireless NoC architecture based on wireless broadcasting through low-power high-bandwidth antennas for redundant and long distant data transmission while employing wired network for local packet delivery.