InfiniBand is a switched-fabric interconnect used across most large HPC and AI training clusters, designed from the start for low latency and RDMA rather than adapted from Ethernet. Host channel adapters support RDMA as a first-class operation, bypassing the kernel entirely and achieving sub-microsecond latencies.

InfiniBand deployments typically use fat-tree topology with adaptive routing to maintain high bisection bandwidth and avoid hotspots.

InfiniBand's native RDMA eliminates kernel transitions.

Conventional TCP:  app -> kernel -> copy -> NIC -> wire
InfiniBand RDMA:   app memory -> HCA -> wire -> HCA -> app memory