Photonic Network-on-Chip (NoC) Architectures for the High Performance Computing Systems

摘要

Higher memory-bandwidth requirement in multi-core computing systems to comply with memory requests from large number of cores can be suitably addressed by replacing traditional electrical network-on-chips (ENoCs) with photonic network-on-chips (PNoCs). Further, energy efficiency of these systems can be significantly improved by replacing electronic on-chip interconnects with silicon nanophotonic interconnects which have higher bandwidth and lower latency. However, the ultimate performance of these systems is limited by the static power for laser sources and waveguide propagation losses. LumiNOC is a nanophotonic network-on-chip (PNoC) architecture, which partitions the entire network into multiple subnetworks for better efficiency. It also uses a distributed arbitration scheme and a channel sharing mechanism for data transmission. Laser sources and ring resonators used in this configuration are expected to have matched optical frequencies for reliable operation. But, thermal sensitivity of photonic devices and process parameters variations inherent during manufacturing process always results in frequency mismatch. An adaptive frequency tuning technique can be used to reduce the difference in resonant frequencies among ring resonators, reduce frequency differences for corresponding on-chip lasers and ultimately to reduce the thermal tuning power of LumiNOC structures.