Designing Scalable Hybrid Wireless NoC for GPGPUs

Data communication in GPU systems exhibits asymmetric patterns that create congestion hotspots. Due to their large number of cores and big die sizes, GPUs also demand highly scalable NoC designs. In this work, we propose hybrid NoC architectures that employ on-chip integrated antennas to build overlaid wireless networks on top of conventional metal/dielectric-based networks. We use low-power high-bandwidth wireless links as express channels to transmit long distance packets and use metal links to deliver local packets. The hybrid architecture can effectively alleviate congestion near traffic hotspots and improve the throughput and scalability of GPU NoCs. We propose solutions for design challenges in such hybrid NoC architectures, such as MAC protocol, router microarchitecture, load balancing and deadlock free routing. To efficiently utilize on-chip wireless bandwidth, we also propose a novel scheme that adaptively allocates bandwidth to wireless channels based on their usage needs. Our evaluation results show that for a GPU with 256 cores, the proposed hybrid architecture can improve performance by 2.4 times on average.