Nvidia contributed Vera Rubin rack and compute tray designs to the Open Compute Project on 12 October and, with Cisco, announced plans for a 400Gb/s-per-lane optical module, compared with 200Gb/s-per-lane pluggable optics.
Key points
- Nvidia’s MGX 1.2 specification supports rack deployments of up to 250kW and includes a liquid-cooled 5,000A busbar.
- Nvidia and Cisco say USFP is designed to provide up to 12.8Tb/s of capacity, with 4x the bandwidth density of 200Gb/s-per-lane pluggable modules.
- TE Connectivity and Amphenol are working with the companies on USFP cages, connectors and modules.
Vera Rubin designs enter OCP’s Open DC Initiative
Nvidia made the rack and tray contributions through its open MGX architecture as part of the Open Compute Project’s Open DC Initiative. The company says the designs give system builders a shared starting point for developing Vera Rubin infrastructure. The announcement came at the OCP Global Summit in San Jose.
The MGX 1.2 specification adds blind-mate liquid cooling at rack scale, interoperable interconnect infrastructure and a liquid-cooled 5,000A busbar. It also provides for 19-inch equipment, configurable I/O bays and fanless, liquid-cooled compute trays. Nvidia says the architecture supports rack deployments of up to 250kW.
Those provisions address the physical work of assembling and servicing high-power racks, alongside the compute design. Nvidia says standardised power and cooling interfaces allow partners to build interoperable systems and improve assembly, testing and field service. In a briefing reported by GamesBeat, Joe DeLaere, Nvidia’s senior manager of AI factory solutions, said the cable-free, hose-free and fanless compute tray could cut tray assembly from hours to minutes.
Nvidia also contributed a Vera CPU rack design specification that extends MGX beyond accelerated computing. The company describes the CPU rack as infrastructure for AI-agent tasks including tool calling and sandboxed execution. An earlier AI Affairs report covered CoreWeave making Vera Rubin available as Cognition ran production workloads.
USFP targets 12.8Tb/s within a switch footprint
The Ultra-Dense Small Form Factor Pluggable, or USFP, design is intended for 400Gb/s-per-lane electrical signals. Nvidia and Cisco say it would let operators connect more computing systems without adding network tiers when switches reach capacity. Additional tiers and optical links can increase latency, power consumption and costs, according to the companies.
The first planned USFP transceivers would fit 16 full-duplex lanes into an octal small form factor, or OSFP, sized footprint for switches and SuperNICs. Nvidia and Cisco say the connector design could later accommodate 32 or more lanes. They put overall capacity at up to 12.8Tb/s, with 4x the bandwidth and front-panel bandwidth density and approximately 1.3x the power efficiency of 200Gb/s-per-lane pluggable modules.
USFP uses a two-dimensional connector whose high-speed contacts engage after insertion. Nvidia and Cisco say that arrangement reduces contact wear and makes insertion force more predictable. The design supports power dissipation of up to 100 watts without a separate embedded cooling circuit in the transceiver. That provision is intended to limit the complexity of switch designs using the higher-density modules.
Nvidia and Cisco are working with TE Connectivity and Amphenol on USFP cages, connectors and modules. The shared format is intended to let multiple suppliers develop hardware for 400Gb/s-per-lane fabrics. Cisco describes the collaboration as a plan to develop an open industry specification.
Cisco cites Microsoft and Alibaba P200 deployments
Cisco’s broader networking announcements at the summit included products at different stages of deployment. Cisco said in a company release that its first Silicon One G300-powered systems were shipping to customers and that systems powered by its Silicon One P200 were operating in production. Those P200 systems are separate from the planned USFP transceivers.
Microsoft said it had moved into production on P200-powered systems. Alibaba said it had deployed its Quantum Cat router, using Cisco’s P200 chip alongside Alibaba’s own systems and the SONiC operating system. Alibaba put the routers’ capacity at 51.2 Tbps.
