Improving GPU NoC Power Efficiency through Dynamic Bandwidth Allocation

High throughput in data communication is of great significance for GPU accelerated systems in order to fully exploit thread level parallelism. Different traffic patterns between GPU NoCs and CPU NoCs lead to suboptimal performance in GPU NoCs that directly adapt from CPU NoCs. Moreover, for GPU NoCs, two networks are usually employed to avoid deadlocks between requests and reply messages. Another important feature of GPU NoCs is the unbalanced traffic load between request network and reply network. This feature often causes the reply network to be congested while the request network is idle. Based on these features of GPU NoCs, this paper proposes a technique called Stop Request Network (SRN). SRN works by stopping request network to reduce energy cost when congestion occurs in the reply network. Our evaluation results show that SRN can save power by 10% with negligible performance degradation.