Watch: Meta’s engineers on building network infrastructure for AI

This post details Meta's network infrastructure evolution for AI workloads, including GenAI training and inference. It highlights the transition to GPU-based training with RoCE-based network fabrics and CLOS topologies. Specific contributions include RDMA deployment with RoCEv2, centralized traffic engineering for dynamic load balancing, and network observability tools (ROCET, PARAM, Chakra) for collective communication analysis. The Arcadia simulator is introduced for end-to-end AI system performance simulation, aiding data-driven decisions and future system design.