What Is Leaf-Spine Architecture?_
Leaf-spine is a two-tier network architecture where every leaf switch (connected to servers) connects to every spine switch, providing consistent latency and predictable bandwidth scaling. It is the dominant Ethernet fabric design for GPU data centers. Adding spine switches increases bandwidth; adding leaf switches increases port count.
Technical Details
Leaf-spine (also called Clos network) architecture provides deterministic, low-latency paths between any two endpoints. Every leaf switch has uplinks to every spine switch, meaning any server can reach any other server in exactly two hops (server → leaf → spine → leaf → server). This predictable latency is essential for GPU cluster workloads. Scaling is straightforward: adding more spine switches increases the aggregate bandwidth between leaves, while adding more leaf switches (with corresponding uplinks to all spines) increases the total port count. The oversubscription ratio (downlinks to uplinks on leaf switches) determines the effective bandwidth between racks. For GPU clusters, 1:1 oversubscription (full bisection bandwidth) is preferred.
How Leviathan Systems Works with Leaf-Spine Architecture
Leviathan Systems deploys the physical infrastructure for leaf-spine Ethernet fabrics in GPU data centers, including switch placement, fiber routing, and testing of all inter-switch connections.