Optical Interconnect Design for Distributed GPU and AI Accelerator Clusters

Abstract

The rapid expansion of distributed GPU and artificial intelligence accelerator clusters has turned interconnect design from a supporting infrastructure concern into a primary determinant of system capability. As model training and inference workloads become increasingly communication-intensive, electrical links face growing pressure in bandwidth density, energy per transmitted bit, reach, and synchronization precision. Optical interconnects offer a promising path toward higher radix, lower latency, and more flexible resource pooling, but they also introduce new architectural, operational, and governance challenges. This paper examines optical interconnect design for distributed accelerator clusters from a system-level perspective. It analyzes workload and traffic characteristics, including collective communication, parameter synchronization, bursty training phases, and latency-sensitive inference. It then evaluates architectural options such as wavelength division multiplexing, optical circuit switching, hybrid electrical and optical fabrics, co-packaged optics, and hollow-core fiber links, with attention to topology, switching granularity, control-plane complexity, and failure containment. The discussion extends to deployment models, composable infrastructure, multi-tenant fairness, robustness, sustainability, and policy implications. The central argument is that optical interconnect design cannot be reduced to a photonic component problem. It is a socio-technical design problem in which performance, economics, energy, reliability, and institutional governance interact. Successful clusters will require co-design across accelerators, network interfaces, optical fabrics, orchestration software, procurement practices, and energy policy. The paper concludes with forward-looking directions for research and deployment.