NVIDIA Unveils AI Factory Innovations at OCP Global Summit

NVIDIA is pushing the boundaries of AI infrastructure, as revealed at the OCP Global Summit. The tech giant is collaborating with over 50 partners to introduce next-generation gigawatt AI factories. The focus is on the NVIDIA Vera Rubin NVL144 MGX-generation open architecture rack servers, designed to meet the increasing demands for AI inference, according to NVIDIA.

Innovative Rack Server Technology

The Vera Rubin NVL144 MGX compute tray is a standout feature, offering a modular, liquid-cooled design that promises improved energy efficiency. This architecture supports NVIDIA ConnectX-9 800GB/s networking and NVIDIA Rubin CPX, which are essential for massive-context AI inference. The design allows for easier assembly and serviceability, making it a significant leap in the field.

Furthermore, the architecture is supported by more than 50 MGX system and component partners, allowing for a mix-and-match capability that enhances scalability. These innovations are set to be contributed as open standards to the Open Compute Project (OCP) consortium.

Transition to 800 VDC Data Centers

NVIDIA is also leading the transition to 800-volt direct current (VDC) data centers, providing increased scalability and energy efficiency over traditional systems. Partners like Foxconn, CoreWeave, and Oracle Cloud Infrastructure are contributing to this shift, with Foxconn unveiling its 40-megawatt Kaohsiung-1 data center in Taiwan.

Vertiv has introduced an efficient 800 VDC MGX reference architecture, while HPE and other partners are announcing product support for the NVIDIA Kyber architecture. This transition mirrors advancements seen in the electric vehicle and solar industries.

NVIDIA Kyber and NVLink Fusion Ecosystem

The NVIDIA Kyber rack server generation is set to succeed the NVIDIA Oberon, featuring innovations in power delivery and cooling. It will integrate 576 NVIDIA Rubin Ultra GPUs by 2027, enhancing AI infrastructure capabilities. The move to 800 VDC architecture promises significant power transmission improvements, reducing copper usage and costs.

NVIDIA’s NVLink Fusion ecosystem is expanding, with Intel and Samsung Foundry joining to integrate custom silicon into NVIDIA’s optimized data center architecture. This collaboration aims to accelerate AI factory scalability and performance.

Open Ecosystem and Industry Collaboration

Over 20 partners are contributing to the development of open-standard rack servers for future gigawatt AI factories. Key players include silicon providers like Analog Devices and Texas Instruments, as well as power system component providers such as Flex and GE Vernova.

The Open Compute Project, founded by Meta, continues to play a crucial role in redefining hardware technology to meet the growing demands of AI compute infrastructure. NVIDIA’s contributions to OCP standards are pivotal in this ongoing evolution.

Image source: Shutterstock