Run-time communication bypassing for energy-efficient, low-latency per-core DVFS on Network-on-Chip

System-level exploration of a novel Network-on-Chip (NoC) architecture with run-time communication bypassing is presented. Fine-grained DVFS (Dynamic Voltage and Frequency Scaling) is an effective power reduction technique. We propose run-time reconfigurable interconnect on each inter-router channel to minimize the latency and energy overhead. When two routers are running on the same frequency, FIFO-channel is bypassed by direct interconnect. Distributed algorithm is designed for per-core DVFS. Proper power delivery and clocking scheme are integrated. Simulation shows significant energy and latency saving.