All-to-All, Low Latency, and High-Throughput Optical Interconnect Demonstrator for Scalable Supercomputing

I. Project Summary This project addresses five key elements for future high-performance computing: (1) high-throughput, (2) lowlatency, (3) energy-efficiency, (4) scalability, and (5) high-radix, all-to-all connectivity. Computing and data centers today are dissipating Megawatts of power, and scaling to an exascale system would require nearly a Gigawatt [1]. Amdahl’s law suggests that a system with balanced computation, memory, and communications performs best under most circumstances, but today’s computing systems are misbalanced by typically more than two orders of magnitude lacking communication bandwidths. Further, they are also lacking high-radix and all-to-all interconnect capabilities that can benefit many important applications. Future exascale high-performance computing and data systems are limited in its scalability and performance due to interconnects, and transformative and new interconnect technologies are essential. This project demonstrates a new hardware and software solutions to interconnects for future high-performance computing. It demonstrates a new computing system with wavelength routed optical interconnects that (1) scale to ~ million compute nodes and Exabyte/s bisection-bandwidth, (2) support all-to-all connectivity without contention or arbitration with the radix count 512 and beyond, (3) exploit ~ pJ/bit communication energy efficiency, (4) achieve normalized throughput of ~100% in Racks, ~97% in Clusters, and ~80% in Full Systems 100% at 100% data injection rate on every transmitter port in the system, (5) validate the performance through an actual hardware system demonstration and a clock-cycle accurate simulator, and (6) demonstrate a chip-scale implementation using silicon photonics.