NVIDIA and Cisco plan USFP optics standard for denser AI networks

NVIDIA and Cisco announced plans for a pluggable optical-module format designed to carry 400 gigabits per second on each electrical lane in AI Ethernet networks. The design, called Ultra-Dense Small Form Factor Pluggable, or USFP, is intended to fit more bandwidth into the front-panel space occupied by current pluggable modules.

Introduced at the OCP event on Oct. 12, USFP is also slated to become an open industry specification through a forthcoming community-led multi-source project. NVIDIA and Cisco have not released the specification, named the project or set a publication date.

The companies say the design would allow network operators to connect more 1.6-terabit-per-second endpoints without adding network tiers. They project four times the bandwidth and front-panel bandwidth density of 200Gb/s-per-lane pluggable modules, along with about 1.3 times the power efficiency. Those figures are company projections, not independently tested product results.

The proposal addresses the network layer around AI systems rather than the compute platforms themselves, such as in Sea’s planned adoption of NVIDIA Vera Rubin.

The initial transceiver design would put 16 full-duplex lanes in the footprint of an OSFP module for switches and SuperNICs. Each lane carries data in both directions. Cisco says 16 lanes running at 400Gb/s per lane would provide 6.4Tb/s of aggregate capacity. It also describes a later path to 32 or more lanes and 12.8Tb/s. The companies have not provided a product availability timetable.

USFP’s connector uses a two-dimensional contact array and a two-step insertion process. Its high-speed contacts engage only after the module is inserted. NVIDIA and Cisco say that approach should make insertion force more predictable, reduce contact wear and improve signal integrity as lane counts rise.

The format is also designed to dissipate as much as 100 watts without a separate embedded cooling circuit inside the transceiver, according to the companies. That thermal capability, like the density and efficiency comparisons, has not been independently demonstrated in deployed hardware.

NVIDIA and Cisco said they are working with TE Connectivity and Amphenol on cages, connectors and modules, and expect additional partners to join.