Solution · AI Data Center

MPO Cabling for GPU Clusters

How to count fibers and MPO links for the backend, frontend and storage networks of a GPU cluster, with a worked example.

Three networks, very different fiber counts

NetworkWhat it carriesFiber demand
Backend (compute fabric)GPU-to-GPU traffic for trainingHighest: usually one high-speed port per GPU
FrontendManagement, users and data ingestLower: a few ports per server
StorageAccess to shared storageMedium: depends on storage design

The backend fabric drives the cabling plan. It is often built rail-optimized, with every GPU in a server connected to a different leaf switch.

Worked example: counting backend fibers

Assumptions for this example only: servers with 8 GPUs, one 400G port per GPU, and 400G DR4 optics (8 fibers per link on MPO-12).

StepCalculationResult
Backend links per server8 GPUs × 1 port8 links
Fibers per server8 links × 8 fibers64 fibers
32 servers32 × 642,048 fibers to the leaf layer
Leaf to spineSame again if the fabric is non-blockingabout 2,048 more fibers

With 800G optics or more GPUs per server, the numbers grow quickly. That is why AI halls use high-fiber-count trunks and dense cassettes rather than individual patch cords.

Send us your server count, GPUs per server, optics and switch layout, and we work out the trunk and cassette quantities in the quote.

Cabling practices for GPU rows

Pre-terminated trunks

Factory-terminated MPO trunks avoid on-site splicing and speed up go-live.

Cassettes at both ends

Cassettes or MPO panels let you re-patch without touching the trunk.

Label every port

Port-level labels on both ends make rail-optimized fabrics manageable.

Spare capacity

Plan spare trunks for upgrades to faster optics and replacement links.

Related guides: Optical interconnects for GPU clusters · Sourcing MPO trunks for AI deployments · 9-point MPO trunk checklist

← Back to AI data center solution

Get a free spec sheet & quote

Tell us the fiber count, length and connector. Candy, our fiber specialist, reviews every request.