OCP Summit 2024: The open future of networking hardware for AI

This post details Meta's contributions to open networking hardware for AI training clusters, including two new disaggregated network fabrics (DSF) and a new NIC (FBNIC). It describes the architecture of DSF, which uses VoQ-based scheduling and OCP-SAI/FBOSS, and highlights specific hardware platforms like Arista 7700R4 series and Minipack3/Cisco 8501 switches. The FBNIC is presented as a multi-host foundational NIC with features like datapath isolation and OCP NIC 3.0 compliance. The post emphasizes the benefits of open hardware for flexibility, scalability, and efficiency in AI infrastructure.