Photonic Interconnect for Next-Generation AI Systems

This conference talk by Ben Lee, Principal Research Scientist at NVIDIA, explores how photonic interconnect technologies can address the challenges in scaling GPU-based AI systems. Learn about the limitations of current electrical signaling in GPU clusters that power AI computing, where cable reach constrains cluster size and creates thermal density issues. Discover how photonic transceivers offer potential solutions through distance-insensitive communication, but must meet strict efficiency, cost, and density requirements to be viable. The presentation examines various link architectures, integration strategies, and optical technologies that could transform intra-cluster fabrics to meet next-generation AI computing demands, ultimately enabling larger and more thermally efficient GPU clusters.