부산대학교 도서관

As the backbone for many-core chips, Network-on-chips (NoCs) consume a significant share of total chip power. As a result, decreasing the power consumption in these components can reduce the total chip's power significantly. NoC's routers can be powered down using power-gating, a promising technique for reducing static power consumption. In some advanced methods, routers are put in sleep mode and only wake up when they are needed to turn/inject packets. Since waking up the router takes several cycles to complete, packets will experience high latency. In this regard, application mapping significantly impacts the number of turns. This article proposes a reinforcement learning (RL) framework based on Actor-Critic architecture to optimize the application mapping problem to minimize the number of turn packets as well as communication cost. Our RL framework learns the heuristic of the mapping problem and outputs a near-optimal mapping. A 2-opt local search algorithm fine-tunes this strategy and provides an improved mapping. Our simulations show that the proposed RL framework can achieve better cost and algorithm run-time performance compared to other heuristic algorithms such as Simulated Annealing (SA) and Genetic Algorithm (GA).