Bandwidth Adaptive Nanophotonic Crossbars with Clockwise/Counter-clockwise Optical Routing

Abstract

Future processors are anticipated to have hundreds or even thousands of processing cores placed entirely on a single silicon chip. The increasing number of cores placed on a single chip presents new challenges, pushing researchers to explore opportunities in emerging technologies such as on-chip silicon nanophotonics. Implications of nanophotonic technology has created a unique landscape for new interconnect designs. Among the many architectures made possible by nanophotonics, there has been notable interest in crossbar topologies that were previously impractical using only electrical components. In this paper, we present a new nanophotonic crossbar interconnect architecture with the aim of retaining the low latency, single-hop characteristic of the crossbar topology, while also improving the networks utility of the static laser source which is often wasted to insertion losses and unused bandwidth. We compare our architecture design to other proposed architectures according to area, power consumption, throughput, and latency. Approximately a 13% improvement in throughput is achieved compared to other optical crossbar topologies and a 92% improvement is achieved compared to a conventional electrical flattened butterfly topology on synthetic traffic patterns.