Nvidia Brings Meta and Google Into New AI Networking Push

Nvidia is expanding its AI infrastructure relationship with Meta and Google through new networking hardware and transport support aimed at large-scale data centers.

Meta is adopting Nvidia’s Spectrum-6 Ethernet switches, which provide 102.4 terabits per second of switching capacity. Google plans to support the Falcon transport protocol on future Nvidia ConnectX-10 SuperNICs and BlueField data processing units.

The announcements were made Oct. 12 at the Open Compute Project Global Summit in San Jose, alongside Nvidia’s contribution of Vera Rubin rack and compute-tray designs through its MGX architecture.

Meta’s N-IceCube switch is designed for the company’s liquid-cooled IcePack rack. Spectrum-6 provides 200 gigabits per second per lane and supports Nvidia’s Multiplane technology across as many as eight network planes. The architecture is intended for AI factories containing more than 100,000 graphics processing units.

“We are excited about our close partnership with NVIDIA in making multiplane fabrics operationally scalable and performant, a critical requirement both for our own infrastructure builds and for our joint cloud partners,” Omar Baldonado, Meta’s senior director of data center and AI networking, said.

Baldonado said Meta’s FBOSS software stack can run across Spectrum application-specific integrated circuits, including the N-IceCube switch. Meta announced an expanded multiyear infrastructure partnership with Nvidia in February and said Spectrum-X would be used across its infrastructure footprint.

Falcon is designed for reliable, low-latency data transfers over Ethernet. Its capabilities include multipath load balancing, congestion control and hardware packet recovery. The Google-related announcement covers support on future ConnectX-10 SuperNICs and BlueField DPUs.

Nvidia’s updated MGX 1.2 specification supports rack deployments of up to 250 kilowatts and includes a liquid-cooled 5,000-amp busbar. Nvidia and Cisco also introduced the Ultra-Dense Small Form Factor Pluggable optical module, targeting 400 gigabits per second per lane with 16 full-duplex lanes and plans for 32.

The developments extend Nvidia’s role in the networking and interconnect systems used to build large AI infrastructure. They follow Nvidia’s presentation of its Vera Rubin systems at the summit, detailed in earlier coverage of the company’s modular AI-factory designs.