NVIDIA And Cisco Reveal USFP Optical Module For AI Factories
The module moves the bottleneck conversation from raw switch capacity to how many links a single switch face can carry, which is why NVIDIA and Cisco want the spec opened to other suppliers rather than kept proprietary.
Reporting from 1 source: ASCII.jp.
NVIDIA and Cisco announced USFP, a pluggable optical module form factor for AI factories, unveiled in Santa Clara, California. The module is designed to support 400Gb/s electrical signaling per lane and reach up to 12.8Tb/s of capacity, with bandwidth density said to improve 4x and power efficiency about 1.3x over conventional 200Gb/s-per-lane pluggable modules. TE Connectivity and Amphenol are expected to join the multi-source ecosystem.
The form factor mounts 16 full-duplex lanes in the same footprint as OSFP, so a switch can carry more connections without adding network tiers. It supports both switches and SuperNICs, with expansion to 32 lanes or more envisioned.
A two-dimensional array connector and a two-stage insertion method are used to improve signal quality and cut contact wear. The design handles heat dissipation of up to 100 watts without a dedicated cooling circuit, which reduces switch design complexity. NVIDIA and Cisco plan to formulate an open industry specification through a community-led multi-source project.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.