Spine-leaf topology for data centers: why three-tier is no longer designed

The three-tier topology (access, aggregation, core) served for two decades a dominant traffic pattern: north-south, from the user to the data center and back. When that pattern stopped being the main one, the topology stopped being the right one. Spine-leaf is not a fad: it is the required redesign when east-west traffic dominates your network.

This article does not compare switch vendors. It explains why three-layer physics and today’s traffic math are incompatible, and why spine-leaf is what hyperscale operators, carriers, and AI data centers in production are doing today.

Why three-tier stopped scaling

In a classic three-tier, two servers in the same rack talk to each other by passing through access to aggregation to core to aggregation to access. Four hops minimum. Each hop adds latency, occupied ports, and a chance of failure. For east-west traffic between workloads in the same rack, this design imposes a fixed cost you can no longer afford.

When inference, training, and microservices workloads exchange traffic among themselves inside the same data center, several times per second, the four hops become a bottleneck. Aggregation saturates, latency grows, and the typical 1:4 to 1:8 oversubscribe ratio becomes visible.

What changes in spine-leaf

Spine-leaf has two layers: leaf (where you connect servers, racks, ToR) and spine (where you connect leaves to each other). Every leaf uplinks to every spine. A server on leaf X talks to a server on leaf Y in exactly two hops: an ECMP to the appropriate spine and the spine down to the destination leaf. The path is predictable, latency is constant, and bandwidth between any pair of servers is identical.

This is what the industry calls “bisection bandwidth”: the total capacity between two halves of the fabric. In a three-tier, that capacity is lower than the aggregate. In a properly sized spine-leaf, every server talks to every other at the uplink bandwidth.

Why spine-leaf won in AI and high-density compute

AI clusters move large volumes of gradients, weight synchronization, and east-west between GPUs. Three-tier imposes more hops per byte transferred, which in large-model training translates into extra hours per run. The difference between 3 hops and 2 hops per GPU-to-GPU communication is not linear: multiplied by thousands of calls per second, it defines whether your cluster finishes in 6 hours or 18.

Cisco, Arista, Juniper, and NVIDIA Mellanox have published reference designs with spine-leaf as the only recommended topology for new deployments of more than two racks of high-density compute. The discussion is not whether spine-leaf is better: it is how many uplinks each leaf needs based on the target density.

Where three-tier is still valid

For campus networks, for office networks, or for small data centers where most traffic leaves to the user (north-south predominant) and not between servers (east-west), three-tier remains a valid and simpler-to-operate option. Not everything needs spine-leaf.

Where it no longer applies is in any environment where most traffic stays inside the same data center, especially when there is AI training, microservices, or object storage accessed from many workers.

The next step

If you are designing a network from scratch or migrating from three or more layers, size with numbers: number of racks, density per rack, traffic profile (east-west vs north-south), uplinks available per leaf, and target oversubscribe ratio (1:1 is typical in modern production). With those numbers spine-leaf sizes without surprises and the migration is done in phases, without prolonged service cutover.

Sources

[1] Cisco — Modern Data Center Fabrics (white paper, spine-leaf) — https://www.cisco.com/c/en/us/solutions/collateral/data-center-virtualization/data-center-networking/white-paper-c11-737895.html

[2] Arista Networks — Spline and Data Center Design Guide — https://www.arista.com/assets/data/pdf/CampusLeaf-Spine-Spline/Design_Guide/Spline-and-Data-Center-Design-Guide.pdf