The transition to AI has necessitated a massive data center infrastructure transition, primarily around power. It’s not just about building nuclear reactors next to big data centers to pump more juice into the servers.
Industry leaders including Nvidia have addressed the need for AI factories with a power distribution system of 800 volts direct current (VDC) alongside multi-timescale energy storage. In many cases, going to 800 VDC will require transitioning from alternating current, at 415 volts AC. Nvidia outlined that change in a new blog,
A number of big silicon providers often associated with sensors are collaborating with Nvidia in the future data center electrical ecosystem, including Infineon Technologies, Navitas Semiconductor, Analog Devices, Renesas, STMicroelectronics and Texas Instruments. Such companies have long been players in a wide range of power components, sensors and related devices. Data center power system providers in the collaboration include GE Vernova, Hitachi Energy , Schneider Electric and Siemens.
“The transition to 800 VDC can’t be accomplished in a vacuum. It requires urgent, focused and industry-wide collaboration,” Nvidia said, noting that Open Compute Project meeting this week will provide a forum for open standards to provide interoperability and potentially reduce costs. “The industry must align on common voltage ranges, connector interfaces and safety practices for 800 VDC environments.”
Moving to 800 VDC will mean reduced copper and cost, because the same wire gauge can carry 157% more power than 415 VAC. That’s due to the three wire set up for VDC, down from four for AC, as well as fewer conductors and smaller connectors.
Arguably more important is that native DC eliminates AC-do-DC conversion steps normally seen in traditional systems. A streamlined power path will boost efficiency and reduce heat. Nvidia noted that 800 VDC is already embraced by the EV and big scale solar industries to improve efficiency.
The vision Nvidia outlined centralizes all AC-to-DC conversion at the facility level to create a native DC data center where medium-voltage AC is directly converted to 800 VDC by large high capacity power conversion systems. Then the 800 VDC is distributed throughout the data hall to compute racks and the architecture is adjusted to eliminate AC switches, transformers and PDUs. Nvidia is prepping a Kyber rack architecture for its MGX architecture.
Nvidia first announced 800 VDC architecture in May and at that time Infineon noted that a new architecture can create risks for system downtimes. Infineon, in collaboration with Nvidia, has announced a new hot-swap technology that relies on advanced semiconductors made of silicon carbide. Data center operators will be able to exchange server boards in an 800 VDC set up while other servers continue to operate in the same rack, according to Infineon.
Infineon will leverage its Intermediate Bus Converter technology and high-frequency switching based on GaN to speed up development of three- and two-stage conversion from grid to core, the company said.
“There is no AI without power. That’s why we are working with Nvidia on intelligent power systems to meet the power demands of future AI data center while providing a serviceable architecture that reduces system downtimes to a minimum,” said Adam White, division president of Power and & Sensor Systems at Infineon. He said the move to 800 VDC requires a revolutionary approach to power. Infineon, based in Munich, employs more than 58,000 workers focused on semiconductors in power systems and IoT.
Texas Instruments said it is also collaborating with Nvidia to develop power management devices to support 800 VDC architecture, noting that the IT rack power is expected to pass the 1MW mark in the next two to three years. The company makes a 30kW AI server power-supply unit to support AI workloads with a three-phase, three-level flying capacitor power correction converter, as well as a dual phase smart power stage and smart power module. It also produces a Gallium-nitride intermediate bus converter in a small form factor.