AI Deployments are Reshaping Intra-Data Center Fiber and Communications (Melissa Farney)

Overview

The article explains how AI data centre networks split into a conventional Ethernet front-end and a dedicated GPU back-end fabric, and what that means for fiber infrastructure. Back-end fabrics need 400-800 Gbps per GPU with sub-microsecond latency in fat-tree or similar topologies, which pushes single-mode fiber and high-power optics. The article also discusses prefabricated cabling to speed up deployment.

In front-end networks, server-to-leaf links run at 25-50 Gbps and spine links at about 100 Gbps, with mostly north-south traffic. Back-end AI fabrics, by contrast, run at 400 or 800 Gbps per GPU with sub-microsecond latency and east-west traffic in fat-tree or hypercube topologies.

Dedicated back-end networks keep GPU utilisation often above 95 percent; on a large cluster each point of idle time can cost hundreds of thousands of dollars per year. Back-end bandwidth has largely eliminated copper, making single-mode fiber the standard for intra-data centre AI links.

800G optical transceivers add significant power and heat, so cooling must be considered. IEEE 802.3 is the related standard. The article is an industry piece offering no method or measurement detail, and its figures are general. The message for designers is that in AI halls the network should be planned as two separate fabrics, each with its own pathways and fiber counts.

Key points

1. Front-end networks: server-to-leaf 25-50 Gbps and spine links about 100 Gbps, mostly north-south traffic.
2. Back-end AI fabrics: 400 or 800 Gbps per GPU, sub-microsecond latency, east-west traffic in fat-tree or hypercube topologies.
3. Dedicated back-end networks keep GPU utilisation often above 95 percent; each point of idle time can cost hundreds of thousands of dollars per year on a large cluster.
4. Back-end bandwidth has largely eliminated copper, making single-mode fiber the standard for intra-data centre AI links.
5. 800G optical transceivers add significant power and heat, requiring cooling consideration.
6. The article also covers how prefabricated cabling speeds deployment.

Using it in design

• Model two separate network fabrics (front-end and back-end) in AI hall templates with separate pathways and fiber counts.
• Include optics power per port in the rack heat load.
• Size cable pathways and fiber counts on a single-mode fiber assumption.
• Offer prefabricated cabling as an option that shortens deployment time.

Who should read it

Designers and network engineers planning network and cabling infrastructure for AI halls.

Related standards

• IEEE 802.3

Publication details

Authors

Melissa Farney

Publisher

Data Center Frontier

Year

2025

Document language

English

Source type

Expert / community

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation