Rail-optimized network
Beginner
A wiring plan where chip 1 of every server plugs into one switch, chip 2 into another, and so on. Matching chips are then one switch apart.
Novice
In servers with eight accelerators and eight NICs, NIC 0 of every server goes to the same leaf switch (rail 0), NIC 1 to rail 1, and so on. Accelerators with the same position in different servers then share a switch, which suits how training traffic is laid out.
Expert
A rail is the set of GPUs with the same local rank across scale-up domains. Rails turn the dominant same-rank traffic into one-hop traffic and multiply the servers under one leaf group by the GPUs per server; cross-rail traffic either climbs to the spine or first hops over the scale-up link (as NCCL’s PXN does).