Nvidia is set to revolutionize data center power management with its new 800V HVDC architecture, aimed at meeting the increasing power demands of 1 MW server racks and beyond, with an expected rollout by 2027. The current power distribution system, operating at 54V DC, is nearing its limits as servers frequently exceed 200 kilowatts. This escalating electrical demand from sophisticated AI chips necessitates a switch to more robust infrastructure to prevent overheating of electrical circuits.
For context, Nvidia’s high-performance chips like the GB200 NVL72 or GB300 NVL72 currently require around eight power shelves. Utilizing the existing 54V distribution would demand an impractical 64 U of rack space, far surpassing typical server rack capacities. Moreover, the challenge extends to the vast amount of copper required; supplying 1 MW using the conventional system would necessitate a formidable 200 kg copper busbar. This would mean that a single gigawatt-capacity AI data center could require up to 500,000 tons of copper, accounting for nearly half of the projected U.S. copper production in 2024.
In response, Nvidia proposes the innovative 800V HVDC system, which connects directly to the site’s 13.8kV AC power source. This advancement is expected to free up considerable rack space while enhancing overall power transmission efficiency in data centers. The transition will also eliminate the cumbersome stages of converting AC to DC and vice versa, simplifying processes that contribute to inefficiencies.
Notably, the implementation of the 800V HVDC will enable a significant reduction in system current for the same power load, enhancing the overall wattage delivery by up to 85% without requiring conductor upgrades. Nvidia estimates that with a lower current flow, it can decrease copper requirements by 45%. Additionally, utilizing direct current systems helps to eliminate inefficiencies associated with alternating current, including skin effect and reactive power losses.
Nvidia is collaborating with industry leaders like Infineon, Texas Instruments, and Navitas to develop this essential system. They are expected to use wide-bandgap semiconductors such as gallium nitride (GaN) and silicon carbide (SiC) to achieve the necessary high power densities for advanced AI performance. The shift to the 800V HVDC architecture is not only a technical challenge but also a critical step towards improving power efficiency, significantly reducing operational costs as data centers evolve to accommodate greater capacities.
This innovation marks a significant stride towards more sustainable and efficient data centers, potentially allowing for cost savings and reduced resource usage in an industry that continues to grow rapidly.