What Is Rail Switch?_
In GPU data-center networking, a rail switch aggregates traffic from a specific 'rail' or lane of the network, often corresponding to a single NIC per GPU node. It is part of a non-blocking fat-tree topology where multiple rail switches work in parallel to provide full bisection bandwidth between GPUs across racks. Rail switches are distinct from leaf or spine switches in that they are dedicated to a single network path per GPU.
Technical Details
Rail switches are typically deployed in pairs per rack for redundancy and to support the dual-port NICs common in H100/B200 nodes. They use QSFP or OSFP ports for 400G/800G Ethernet or NDR InfiniBand connections, with fiber optic cabling to spine switches. The switch ASIC must support RoCEv2 or InfiniBand transport for low-latency GPU-to-GPU communication. Cable management and port mapping are critical to avoid cross-rail congestion.
How Leviathan Systems Works with Rail Switch
During Leviathan Systems rack builds, rail switches are mounted in the top-of-rack position and cabled to each GPU node's NIC ports using fiber MPO trunks. Our field crew verifies that each rail switch's uplinks to the spine fabric are properly labeled and routed per the customer's cabling diagram.
Related service
Network Testing & Commissioning →