IT News on October 12th: NVIDIA announced yesterday (October 12th) that at the Open Compute Project (OCP) Global Summit, NVIDIA disclosed the design for Vera Rubin racks and computing brackets, and together with Cisco, Google, Meta, Microsoft, Visionbay, and others, they released several AI infrastructure open standards.

At this OCP Global Summit, NVIDIA announced that by making its NVIDIA MGX architecture open, it will contribute Vera Rubin rack and compute bracket designs to Open Compute Project ( OCP ). System manufacturers can use these designs to develop infrastructure for the next generation of AI factories.

The updated MGX 1.2 specification includes new rack-level blind-mate liquid cooling, interoperable rack interconnections, liquid-cooled 5000A busbars, and support for native 19-inch equipment. As cited in a blog post from IT, this architecture supports rack deployments of up to 250kW and simplifies the assembly, testing, maintenance, and deployment processes.

NVIDIA also released 4 open-source software components of DSX AI Factory, covering switch configuration, GPU cluster stress testing, rack firmware management, and bare-metal deployment. These components are derived from NVIDIA DGX Cloud operational experience and can help cloud service providers deploy hardware into production clusters within a few hours.

In terms of cooperation, NVIDIA also announced that it has collaborated with AMD, Google, and Meta to release Open Data Center Specification v0.7, unifying the power supply, cooling, and physical layout design for the AI factories.

Google, Meta, Microsoft and NVIDIA have additionally issued grid resilience specifications, requiring large-scale AI facilities to maintain connectivity during grid disturbances and to recover to near full load within less than 2 seconds.

Meta will adopt NVIDIA Spectrum -6 102.4Tb/s Ethernet switches to build a Vera Rubin AI factory based on an open network. The switching capacity and single-channel bandwidth of Spectrum-6 are both twice that of Spectrum-4. Combined with Spectrum - X Multiplane technology, it can support more than 100,000 clusters of GPU.

NVIDIA and Cisco have released the USFP standard, which is the industry's first pluggable optical module designed for 400Gb/s network architectures. USFP provides 16 full-duplex channels within a size similar to that of traditional 200Gb/s single-channel pluggable optical modules, offering a bandwidth density four times that of the latter, and an energy efficiency improvement of approximately 1.3 times.

NVIDIA and Google announced that in the future, ConnectX-10, SuperNIC, and BlueField DPU will support Falcon Transport to enhance the reliability and bandwidth utilization efficiency of Ethernet AI networks.