Abstract
General-purpose computing on graphics processing units (GPGPU) can provide orders of magnitude more computing power than general purpose processors (CPU) for highly parallel applications. For such parallel applications, the memory traffic pattern of GPGPUs behaves considerably different from that of CPUs. This gives rise to opportunities for optimizing the on-chip interconnection network (NoC) of GPGPUs. In this work, we first investigate the characteristics of GPGPU memory traffic of typical benchmarks and categorize the memory traffic patterns. Different traffic patterns require different throughput in the request and reply paths of the NoC to match the network load. To meet this requirement, we examine the feasibility of scaling the network frequency dynamically to balance the throughput of the request and reply networks. The decision is guided by monitoring some shader cores to identify the memory traffic pattern. Performance evaluation shows that this dynamic frequency tuning design can achieve up to 27% improvement in terms of execution speedup compared to a baseline setting and 7.4% improvement on average.