Global Tech News For people figuring out what to build with AI.
AI

NVIDIA and Cisco plan 6.4 Tb/s pluggable optics for AI networks

A disassembled 40-gigabit-per-second QSFP optical transceiver showing its electronics and fiber interface; this is an earlier module, not the announced USFP design.

NVIDIA and Cisco say a planned optical module for AI data centers will carry 6.4 terabits per second (Tb/s) in its first configuration, four times the bandwidth density of today's 1.6-Tb/s pluggable optics in a similar footprint. Announced at the Open Compute Project Global Summit on October 12, the Ultra-Dense Small Form Factor Pluggable (USFP) design uses 16 lanes rated at 400 gigabits per second each. The companies describe a later 12.8-Tb/s version as a possible 32-lane expansion, not the first module.

More data lanes in a familiar module footprint

AI accelerators exchange data across networks while training and serving models. Optical transceivers convert switch or network-card signals into light for fiber links, then back into electrical signals at the other end. That makes the module's bandwidth and the number of modules a switch can accommodate relevant to how large a cluster can grow.

The first USFP transceivers are planned for an OSFP-sized footprint and are intended for switches and SuperNICs, network adapters designed for high-throughput servers. NVIDIA and Cisco say 16 full-duplex lanes at 400 Gb/s each provide 6.4 Tb/s of aggregate capacity. They compare that with 1.6-Tb/s modules using 200-Gb/s lanes and claim four times the front-panel bandwidth density. The comparison is a design specification, not a result from an independently tested deployment.

Future capacity, power and ecosystem plans

The companies say the connector could support 32 full-duplex lanes in a future configuration, which would raise total capacity to 12.8 Tb/s. They also claim about 1.3 times the power efficiency of 200-Gb/s-per-lane pluggable modules, but the announcement does not provide a measurement method or operating conditions.

NVIDIA and Cisco are working with TE Connectivity and Amphenol on cages, connectors and modules. They plan to propose an open industry specification through a community-led, multi-source project. The announcement describes design plans; it gives no shipping date or customer deployment.

Sources

Illustrative image: Dragon Sully, “Disassembled QSFP transceiver,” Wikimedia Commons, CC0 1.0. It shows an earlier 40-Gb/s module, not USFP.