Fat-Tree Topology

A fat-tree is a multi-layer network topology in which the number of links (and therefore bandwidth) increases towards the core of the network, ensuring that any-to-any communication can occur at full line rate — no bandwidth bottleneck exists at the core. This property, called full bisection bandwidth, is essential for distributed AI training where every GPU must communicate with every other GPU at high speed.

In practice, fat-tree networks are implemented using a multi-stage Clos design. A k-port switch fat-tree can connect k³/4 hosts with k² core switches and k²/2 aggregation switches. InfiniBand HDR/NDR fabrics for AI clusters commonly use fat-tree or flattened butterfly topologies. Spine-leaf is a simplified two-tier fat-tree, appropriate when full bisection bandwidth is not required.