Rail-optimized network

A rail-optimized network is a GPU cluster networking topology that connects the same-indexed NIC on every server to the same leaf switch, creating dedicated "rails" so multi-node GPU-to-GPU traffic reaches its destination in the fewest possible hops. This reduces bandwidth contention and latency compared with generic top-of-rack designs, and is the reference architecture behind Nvidia's DGX SuperPOD and large hyperscaler AI clusters. Rail-optimized fabrics typically run over InfiniBand or RoCE Ethernet and pair with NVLink's intra-server GPU interconnect, forming a two-tier backbone — intra-server NVLink plus inter-server rail-optimized fabric — that lets thousands of GPUs act as a single training cluster. The design materially affects a facility's power and cabling density requirements, feeding directly into colocation pricing for AI-ready space.