Spine-leaf
A network layout where every server switch connects to every core switch, so any two servers are only a few hops apart.
Also: leaf-spine, fat tree, Clos network
In a spine-leaf network, servers plug into leaf switches, and every leaf connects to every spine switch above it. Traffic between any two servers crosses at most a leaf, a spine and another leaf, giving even delay and many parallel paths. Large AI clusters add a third layer, often called the super-spine, to reach tens of thousands of GPUs. When every layer has as much capacity going up as coming in, the network is called non-blocking or a full fat tree, which AI training usually needs.