Nvidia introduced new open-source contributions at the Open Compute Project Global Summit, sharing Vera Rubin rack and compute tray designs via its MGX architecture. These blueprints aim to standardize the development of AI factories, including rack layouts, power, and cooling systems.
The MGX 1.2 update introduces modular components for liquid-cooled, fanless I/O bays. Nvidia claims its new Vera CPUs offer twice the performance of the previous Grace generation across several benchmarks and a 1.8-times improvement on Terminal-Bench.
Nvidia is collaborating with Meta, Google, and Cisco to enhance networking infrastructure. Meta will adopt the 102.4 terabit-per-second Spectrum-6 Ethernet switch for its next-generation AI factories to manage massive data traffic.
Nvidia is integrating Google’s Falcon Transport technology into its Spectrum-X ecosystem, starting with future ConnectX-10 SuperNICs. Additionally, Nvidia and Cisco are developing the USFP optical module format to support 400-gigabit-per-second-per-lane networking.
The new USFP design provides four times the bandwidth density of current 200G OSFP modules and improves power efficiency by 1.3 times. These advancements allow data centers to increase network capacity within existing physical footprints.
