IP Fast Reroute for AI/ML Fabrics

This document describes the requirements and mechanisms for achieving sub-100 microsecond convergence in Artificial Intelligence (AI) Data Center (DC) fabrics and Data Center Interconnect (DCI) environments. It explores the limitations of current IP Fast Reroute (RFC 5714) capabilities, such as ECMP, LFA, and TI-LFA, particularly in the context of large-scale, multi-tier Clos topologies and BGP-only fabrics. The draft highlights the requirements for hardware- accelerated network notification mechanisms and congestion-aware remote protection strategies to address the stringent performance demands of AI workloads.