Nvidia and Cisco Introduce 12.8 Tbit/s USFP Modules for AI Fabrics

Nvidia and Cisco have announced new Ultra-Dense Small Form-Factor Pluggable (USFP) modules, offering speeds of up to 12.8 terabits per second (Tbit/s). These represent the first modules to achieve 400 gigabits per second (Gbit/s) per lane and are targeted at organizations operating AI fabrics.

The collaboration between the technology firms expands upon their previous Spectrum-X reference solution for GPU clusters. The new USFP format is designed to tackle the primary challenges in AI networking: bandwidth requirements, power consumption, and heat dissipation. Manufacturers claim these modules provide 1.3 times higher energy efficiency and four times greater bandwidth density compared to transceivers operating at 200 Gbit/s per lane.

In conjunction with TE Connectivity and Amphenol, Nvidia and Cisco aim to establish a hardware ecosystem. Initial USFP transceivers will be compatible with switches and SuperNICs featuring 16 × 400 Gbit/s full-duplex lanes. These will initially deliver 6.4 Tbit/s per module.

The USFP format supports up to 100 watts of power loss without requiring a dedicated cooling circuit, a feature intended to simplify switch design and enhance reliability. The format is positioned as a scalable solution, supporting both 200G and 400G SerDes-based systems and ASICs, enabling an upgrade path from current 1.6 Tbit/s fabrics to 12.8 Tbit/s.