AI Fabric Sizing Calculator

Given a GPU deployment target: leaf/spine switch counts, explicit port accounting, an optics and cable BOM, and a fabric power estimate.

Scope — scale-out fabric only. This tool sizes the backend scale-out fabric (leaf/spine RoCE Ethernet or InfiniBand between servers). The scale-up/intra-node interconnect (NVLink, Infinity Fabric, UALink, Scale-Up Ethernet) is an input assumption — it determines NICs per server via the platform preset — and is never itself sized here.

Try a worked example:1,024 × H100, rail-optimizedsame cluster, spineless (rail-only)256 × MI300X8,192 GPUs on Tomahawk 616,384 GPUs at the 2-tier limit512-GPU disaggregated inference

How it's computed — the conventions behind the numbers
  • Leaf split: downlinks are maximized against the cage budget — ceil(D/f_down) + ceil(U/f_up) ≤ cages, where breakout factors let 800G/1.6T cages serve finer-grained links on either side.
  • Spine count: rounds up to the smallest divisor of uplinks-per-leaf (vendor practice — even ECMP); the honest capacity minimum is available as a mode and warns when ECMP would be uneven.
  • Rail placement: small clusters pack multiple complete rails per switch, matching published vendor designs; strict one-rail-per-switch is available as a mode.
  • Optics: 2 transceivers per fiber link (both ends counted), 1 cable per DAC link, spares rounded up per line. Media is chosen by reach class; if no sourced record matches, the tool warns instead of guessing.
  • Power: switch typical draw (excludes-optics convention, so nothing is double-counted) + per-end optics draw; spares draw nothing.
  • Ceilings: 2-tier designs require a full leaf–spine mesh, so the max is spine-radix bound; the tool warns at 80% and refuses to silently clamp beyond it.

Every formula is documented, human-reviewed, and locked by golden tests that reproduce the published reference architectures listed below. Divergences are documented, never silently adjusted.

Feedback

Does this match what you'd build — or what a vendor quoted you? Both answers are useful.Send feedback(your current design's link is included), or useCopy shareable link above to pass the exact design to a colleague.

Validated against

The calculation engine reproduces the switch and cable counts of these published documents (pinned revisions) in its automated test suite. Divergences are documented, never silently adjusted.