AI fabric sizing tool
Size a non-blocking leaf-spine fabric for a GPU cluster. Enter your GPU count, port speed, and switch silicon to see how many leaf and spine switches the fabric needs, the fabric link count, and the oversubscription ratio. This sizes the network only; it is not a cost estimate.
…
Counts assume a folded two-tier leaf-spine (Clos) fabric with un NIC fabric per GPU, where each leaf splits its ports between GPUs and spine uplinks. Clusters above the pod ceiling need multiple pods or a three-tier fabric, and the tool flags this and shows the pod count. Real AI fabrics add rail-optimized wiring (multiple NICs per GPU for traffic locality) and grow into a 3-stage Clos at scale; the leaf and spine counts match these reference designs. See the Progetti di riferimento della topologia AI fabric.
Tomahawk platforms for AI fabric
The fabric above runs a single OcNOS-DC image on open Broadcom Tomahawk hardware. These are the supported 800G and 400G platforms it sizes onto, from Edgecore and UfiSpace.
How the sizing works
The tool models a folded two-tier leaf-spine fabric, the standard topology for RDMA and AI workloads, and reports the structural switch and link counts. It does not estimate cost.
- Switch radix is the number of ports a switch presents at the chosen GPU speed: Tomahawk 5 gives 64 ports at 800G or 128 at 400G; Tomahawk 4 gives 64 ports at 400G. (Tomahawk 4 uses 50G-PAM4 SerDes, so it has no native 800G; 800G needs Tomahawk 5.)
- Leaf split: a non-blocking (1:1) leaf dedicates half its ports to GPUs and half to spine uplinks; a 3:1 leaf dedicates three quarters to GPUs.
- Leaf count = GPUs ÷ GPU-facing ports per leaf. Spine count = the fewest spines that can carry every leaf's uplinks non-blocking, growing with the cluster up to the uplinks-per-leaf at a full pod (a single-leaf cluster needs no spine).
- Pod ceiling = the largest GPU count a single two-tier pod can hold at that radix; above it, the design needs multiple pods or a super-spine tier.
This tool provides a structural network-sizing estimate for planning only and is not a performance guarantee, a bill of materials, or a cost estimate. Actual designs depend on rail optimization, NIC count per GPU, cabling, and failure-domain choices. Broadcom and Tomahawk are trademarks of Broadcom Inc.; other names are the trademarks of their respective owners. Talk to IP Infusion for a design reviewed against your specific platforms.