Evolution of Ethernet Based Switch Platforms and Fabrics to Meet Meta's AI Training Clusters

Learn how Meta evolved Ethernet-based switch platforms and fabrics to support large-scale AI training clusters in this 23-minute conference talk from OCP 2024. Discover the journey from deep-buffered to shallow-buffer disaggregated Ethernet platforms for building non-blocking fabrics that interconnect GPU clusters of up to 4,000 units. Explore the switching and routing platforms, features, and unique scaling challenges encountered when constructing significantly larger non-blocking fabrics for generative AI training workloads. Understand the specific SAI and FBOSS enhancements that enabled adaptation to new platform features, including the evolution from cell-based fabrics to more scalable architectures. Gain insights into the technical decisions and engineering solutions that Meta's software and hardware engineers implemented to meet the demanding requirements of modern AI training infrastructure.