AI Fabric vs Traditional CLOS
What stays the same from your spine-leaf experience and what's different in an AI fabric — 1:1 oversub, rail-optimized topology, RDMA-aware QoS, and zero tolerance for tail latency.
Rail-Optimized Topology
What "rails" mean in an AI fabric, why each GPU gets its own dedicated leaf, how this changes blast radius, and pod sizing.
Hash Polarization & Elephant Flows
Why traditional ECMP balances web traffic but breaks under synchronized AI training collectives. The elephant-flow problem, hash polarization, and what hyperscalers do about it.