Energy-Efficient and High-Performance Nanophotonic Interconnects for Shared Memory Multicores

By 2018, both industry and academic leaders expect computer performance to reach the exascale level, allowing 10^18 floating point calculations persecond (FLOP). However, exascale computing can only exist if high-performance and energy-efficient chip multi-processors (CMPs) can be realized. It is well-known that the power dissipation of metallic interconnects in future multicore architectures is projected to be a major bottleneck as we scale to the sub-nanometer regime. This has forced researchers to focus their attention on developing alternate power-efficient technology solutions for performance limitations of future multicore architectures. Nanophotonic interconnects are a disruptive technology solution that is capable of delivering the communication bandwidth at low power dissipation when the number of cores is scaled to large numbers. Furthermore, recent advances in complementary metal-oxide semiconductor (CMOS) compatible devices and circuits development have made nanophotonics a viable solution for on-chip applications. This dissertation proposes two nanophotonic architectures called 3D-NoC and PULSE. These architectures alleviate two major problems facing current mutlicores, which are to improve performance and programmability respectively. 3D-NoC combines the benefits of 3D integration with nanophotonics to construct a high-performance high-core (256 cores) CMP. In addition, to further maximize the performance of 3D-NoC, this dissertation proposes a reconfiguration algorithm whose purpose is to improve performance by adapting available network bandwidth to application demands. This is accomplished by monitoring the traffic load and applying a reconfiguration algorithm that works in the background without disrupting the on-going communication. PULSE is a tree-based broadcast network which combines/splits optical signals using a combination of couplers and splitters in such a way that the same intensity light arrives at all the cores simultaneously ensuring the ordering required for snoopy protocols. In addition, this dissertation proposes a photonic cache filtering technique called multi-PULSE. This allows the broadcast network to rapidly morph into a multicast network by directing the address request to only those cores that actually share the block. Multi-PULSE allows for a reduction in cache access power as only the cores that have the cache block will receive the request. Moreover, 3D-NoC and PULSE are compared to other leading electrical and nanophotonic architectures using synthetic traffic, SPEC CPU2006, Splash-2 and PARSEC benchmarks. When PULSE is compared to other leading nanophotonic and electrical broadcast networks, simulation results show PULSE demonstrates a speed-up of 55% and power savings of 80% over other leading networks. On the other hand, when 3D-NoC is compared to other leading nanophotonic networks, simulation results indicate that 3D-NoC can further improve Splash-2, Parsec, and SPEC CPU20006 benchmarks by 10%-25% .