Pod Backend Fabric

Analyzes rail-optimized, condensed, and multi-pod GPU fabric topologies for B300 clusters, comparing latency, cost, and workload tradeoffs across collective and serving patterns.

Reframes GPU cluster topology around pod-level isolation domains, where instance placement and intra-pod fabric design replace cross-pod sharding and global fabric homogeneity.

Introduces collectives, parallelism types, KV cache, and NVMe-oF vocabulary, then explores scheduling complexity when pods are isolation domains and workloads exceed cluster capacity.