KARL: A Cost-effective Routing Algorithm in Fault Tolerant 3D Network-on-Chip via K-means Assisted Reinforcement Learning
Lujian Chen, Jiajia Jiao, Ruirui Shen · 2022
In order to guarantee the communication quality of large-scale multicore processors, 3D Network-on-Chip has become the dominate component for interconnecting various IPs via using 3D stacking technology. A cost-effective routing algorithm KARL is proposed for fault tolerant 3D NoC using K-means to shrink the large Q-table of reinforcement learning with minimal performance loss. The simulation results from gem5 simulator demonstrate that the proposed adaptive routing approach can achieve the average latency 8.26% reduction and successful delivered rate 5.50% improvement over the state of the art on fault tolerant routing algorithms for 3D NoCs. The Q-table can be shrink 70% with less than 10% performance loss.