Beyond GPUs to Data Center Design: Nvidia Opens Up 'Vera Rubin' Infrastructure

Nvidia has opened up the infrastructure design supporting its next-generation AI platform, "Vera Rubin." The move extends beyond chips into data center design and operations, with the goal of laying a common foundation for building AI data centers (AIDC).

The scope of the opening includes server rack and compute tray designs. It also covers power and cooling specifications, as well as operating software.

Nvidia announced the details on the 12th at the Open Compute Project (OCP) Global Summit in San Jose, California. It said it contributed key Vera Rubin rack and compute tray designs to OCP through its open MGX architecture. It also expanded open-source software components for DSX, its AI infrastructure operations platform. The direction of the announcement is to design not only around GPU performance, but also around facilities and operating systems.

Earlier, Nvidia unveiled the Vera Rubin DSX reference design in March. The reference design outlined how to integrate computing, networking, storage, power, and cooling systems. Nvidia also introduced the digital-twin-based "Omniverse DSX Blueprint," a tool designed to pre-validate facility layout, heat flow, and power configuration.

The MGX 1.2 configuration for Vera Rubin consists of rack-level liquid cooling, interoperable interconnect infrastructure, and a liquid-cooled 5,000A busbar. The busbar serves as the conductor that delivers power to equipment. MGX 1.2 applies standardized power and cooling interfaces, as well as fanless liquid-cooled compute trays, and supports rack configurations of up to 250kW.

Nvidia said MGX 1.2 standardizes the power, cooling, and connectivity foundation at the rack level, making it possible to build high-density liquid-cooled racks. According to the company, this can shorten assembly, testing, and maintenance time.

At the event, Nvidia also introduced Open Data Center Standard 0.7, jointly developed by AMD, Google, Meta, and Nvidia. The standard contains structural, power, and cooling design guidelines for AI and machine learning systems after 2027-2028, and sets common specifications for rack weight, space allocation, power density, and cooling infrastructure. Nvidia said the goal is to accommodate equipment from a variety of vendors.

The joint standard serves as a design readiness benchmark for equipment makers and data center operators. It is intended to ease the burden of repeatedly creating custom designs to meet customer-specific requirements and to help prepare facilities in advance for next-generation equipment. An OCP document aims to expand the design's applicability beyond in-house data centers to colocation sites and third-party providers.

Against this backdrop, the opening is expanding into software as well. Nvidia added open-source components to DSX OS. The additions cover network configuration, cluster verification, rack management, and physical server deployment.

These include NVIDIA Config Manager and Cluster Readiness Engine. NVIDIA Config Manager is an element that automates switch configuration. Cluster Readiness Engine checks the status of GPU clusters before training begins.

The expanded scope of the opening also extends to power grid connection methods, which are included among the jointly designed items. The OCP standard involving Google, Meta, Microsoft, and Nvidia includes how data centers should respond when voltage disturbances occur. The standard discussions arose amid concerns that sudden swings in the power demand of large facilities could affect grid stability. The standard aims to standardize response criteria between facilities and utilities, reducing the burden of on-site analysis and design changes.

Through DSX, Nvidia has proposed an approach to operate data centers as integrated systems rather than as collections of servers. The scope of integrated operations extends to controlling power and cooling together. Nvidia introduced DSX Flex and DSX Exchange within DSX. DSX Flex has a function that adjusts power use in connection with power grid services, while DSX Exchange connects signals from computing, power, and cooling equipment.

This time, the company included rack design specifications, facility design specifications, and some operating software in the scope of the opening. The focus is on expanding the number of companies developing equipment and facilities suited to Nvidia's platform. The purpose of the opening was also set around broadening touchpoints.

This is seen as part of a broader effort to address the integration of power, cooling, and facilities, a challenge that cannot be solved by GPU supply alone. Nvidia chose to provide common standards, aiming to reduce the complexity of infrastructure deployment as a business strategy. It also seeks to expand the adoption base for Vera Rubin by encouraging participation from equipment makers, facility operators, and software companies.

The open standards could be used by South Korean AIDC-related companies and may serve as a reference benchmark for product development and compatibility validation among cooling, power, rack, and operating software companies.

However, turning the standards into actual business opportunities requires product implementation capabilities and integrated validation capabilities. An industry source said that simply using the open standards does not automatically secure the operational stability customers demand, nor does it automatically secure the level of fault response customers require. The source added that for South Korean companies to win real supply opportunities, they need maintenance capabilities and technical support responsibility beyond equipment delivery.

Source: IT DAILY · Lee Jae-young
Original: https://www.itdaily.kr/news/articleView.html?idxno=242145

References

• https://www.itdaily.kr/news/articleView.html?idxno=242145