A two-tier non-blocking fabric where every leaf connects to every spine — the standard data-centre architecture.
What the Topology Guarantees
A leaf-spine fabric has exactly two roles. Servers attach to leaves; leaves attach to every spine; nothing attaches sideways. Any server reaches any other by going up once and down once, so the path length is identical for every pair — which means latency is predictable and equal-cost multipath has genuinely equivalent choices to make.
It displaced the three-tier access, aggregation and core design because data-centre traffic became predominantly east-west. A tree optimised for traffic leaving the building handles traffic crossing it badly, and AI workloads are almost entirely traffic crossing it.
The Two Numbers That Matter
Oversubscription is downlink capacity divided by uplink capacity. A leaf with 48 downlinks and 12 uplinks at the same rate is 4:1, which is unremarkable for general compute and unacceptable for training, where a collective can saturate every port at once. AI fabrics are built at or near 1:1.
Spine radix sets the ceiling. The number of leaves a single spine tier can serve is the spine’s port count, so a 64-port spine supports 64 leaves and no more. Current merchant silicon reaches 102.4 Tb/s in one chip — 64 ports of 1.6T, 128 of 800G or 512 of 200G — and higher radix at lower port speed is often the more useful trade, because it postpones the third tier.
What It Costs in Fibre
Leaf-to-spine links number leaves multiplied by spines, and every one of them is a structured run between rows. Thirty-two leaves and sixteen spines is 512 links, which is 512 pairs of transceivers and 512 fibre paths through the plant — the single largest line item in most fabric builds.
Because the runs are between rows rather than within racks, they are single-mode territory in any reasonably sized hall, and they are trunked. The economics favour MPO trunks with cassettes at each end over point-to-point jumpers, which is why MPO trunk cable and MPO to LC cassettes dominate this tier. See single-mode vs multimode for where the crossover sits.
Where You Meet It
Essentially every modern data centre, and as the substrate underneath every GPU fabric — a rail-optimised design is a leaf-spine fabric with the leaf assignment changed, not a different topology. Breakout is how the two tiers are matched: an 800G spine-facing port split into two 400G leaf uplinks is routine, and the breakout configurator will tell you which splits the platform actually supports.
Leaf-spine reference
| Also called | folded Clos, or a two-tier fat tree |
|---|---|
| Server-to-server path | leaf, spine, leaf — three switches, one hop up and one down |
| Oversubscription | downlink capacity divided by uplink capacity |
| Non-blocking leaf | uplink bandwidth equals downlink bandwidth |
| Leaf-to-spine links | number of leaves × number of spines |
| Max leaves | set by spine port count (radix) |
| Typical AI ratio | 1:1, or close to it |
| Third tier | added when spine radix runs out |