Switching · 7 min read · 19 November 2026

Leaf-Spine vs Three-Tier Network Architecture: Which Fits Your Data Center?

How leaf-spine and three-tier (core, aggregation, access) networks differ, why data centers moved to leaf-spine, how to size one, and where three-tier still makes sense.

A three-tier network stacks access, aggregation and core switches in a tree. A leaf-spine network connects every leaf switch to every spine switch, so any server reaches any other server in two hops with the same latency and bandwidth. Leaf-spine is the standard for modern data centers and server rooms; three-tier (or a collapsed two-tier core) remains common for campus networks.

The two designs

Three-tierLeaf-spine
LayersAccess, aggregation (distribution), coreLeaf (servers connect here) and spine
Traffic it suitsNorth-south: users to servers and the internetEast-west: server to server, storage, virtualisation, AI
Loop preventionOften spanning tree, with blocked linksRouted fabric with ECMP: all links active
LatencyVaries with the pathPredictable: two hops between any two servers
GrowthAdd aggregation blocks; may need redesignAdd leafs for racks, spines for capacity
Typical homeCampus, buildings, branch networksData centers, server rooms, GPU clusters

Why data centers moved to leaf-spine

Virtualisation, distributed storage and AI made most traffic flow between servers rather than to users. In a tree, that traffic climbs to the core and back, and spanning tree blocks half the links. A leaf-spine fabric routes on every link at once using ECMP, so adding a spine adds usable bandwidth for every rack.

Sizing a fabric in five steps

  1. Count server ports: servers × 2, because each server connects to two leafs.
  2. Choose the leaf speed: 10G, 25G, or 100/200G for GPU servers.
  3. Number of leafs: ports ÷ ports per leaf, rounded up to an even number.
  4. Check oversubscription: server bandwidth ÷ uplink bandwidth per leaf; 3:1 or better, 1:1 for storage and AI.
  5. Spines: at least two, with enough ports for every leaf uplink.

Modern fabrics add VXLAN and EVPN to stretch Layer 2 networks across the routed fabric, and M-LAG so servers use both links without spanning tree.

Where three-tier still fits

For a campus or building network, users and access points connect at the access layer and most traffic goes to the internet or the data center. A resilient core, often a chassis or a pair of Layer 3 switches, with aggregation per building, is simpler and cheaper there. Many organisations run both: a campus core and a leaf-spine fabric in the server room below it.

Navijin offers NJD data-center switches for leaf and spine, from 10G to 400G, and NJC chassis switches for campus cores.

FAQ

Frequently asked questions

How many spine switches do I need?

At least two for redundancy. Add spines when you need more bandwidth or more leaf uplink ports.

Can a small server room use leaf-spine?

A pair of switches acting as both leaf and core is often enough. Move to leaf-spine when you have several racks or heavy east-west traffic.

Does leaf-spine need spanning tree?

No. The fabric is routed and uses ECMP; servers dual-home with M-LAG or LACP to a leaf pair.

Need help designing your network?

Navijin engineers provide free design consultations, bills of materials and quotations for business projects.