Silicon Nanophotonics for Many-Core On-Chip Networks

Number of cores in many-core architectures are scaling to unprecedented levels requiring ever increasing communication capacity. Traditionally, architects follow the path of higher throughput at the expense of latency. This trend has evolved into being problematic for performance in many-core architectures. Moreover, the trends of power consumption is increasing with system scaling mandating nontraditional solutions. Nanophotonics can address these problems, offering benefits in the three frontiers of many-core processor design: Latency, bandwidth, and power. Nanophotonics leverage circuit-switching flow control allowing low latency; in addition, the power consumption of optical links is significantly lower compared to their electrical counterparts at intermediate and long links. Finally, through wave division multiplexing, we can keep the high bandwidth trends without sacrificing the throughput. This thesis focuses on realizing nanophotonics for communication in many-core architectures at different design levels considering reliability challenges that our fabrication and measurements reveal. First, we study how to design on-chip networks for low latency, low power, and high bandwidth by exploiting the full potential of nanophotonics. The design process con- siders device level limitations and capabilities on one hand, and system level demands in terms of power and performance on the other hand. The design involves the choice of devices, designing the optical link, the topology, the arbitration technique, and the routing mechanism. Next, we address the problem of reliability in on-chip networks. Reliability not only degrades performance but can block communication. Hence, we propose a reliability-aware design flow and present a reliability management technique based on this flow to address reliability in the system. In the proposed flow reliability is modeled and analyzed for at the device, architecture, and system level. Our reliability management technique is superior to existing solutions in terms of power and performance. In fact, our solution can scale to thousand core with low overhead