The aim is to focus is on creating a new architecture for power systems using centralised, HVDC distribution, specifically at 800V.
At present, the power supply in AI data centers is decentralized. This means that the AI chips are supplied with power by a large number of PSUs. The new architecture will be centralised, making the best possible use of the constrained space in a server rack.
Most data centres distribute power as AC which is then converted to DC by individual PSUs inside servers. This process leads to significant energy losses during conversion and distribution.
The new approach is to supply power directly to server racks and convert it only at the point of use (the AI chip/GPU). This reduces conversion stages, increases efficiency, and supports higher power densities. The approach will use fewer power conversion stages and allow upgrades to higher distribution voltages.
With AI datacentres expected to require over 1MW per server rack by the end of the decade, the HVDC system is designed to improve energy efficiency, reliability, and scalability for next-generation AI workloads.
This paves the way for the implementation of advanced power delivery architectures in accelerated computing datacentres and will further enhance reliability and efficiency. As AI datacentres already are going beyond 100,000 individual GPUs, the need for more efficient power delivery is becoming increasingly important.
“The new 800V HVDC system architecture delivers high reliability, energy-efficient power distribution across the datacentre,” says Gabriele Gorla, vice president of system engineering at Nvidia, “through this innovative approach, Nvidia is able to optimize the energy consumption of our advanced AI infrastructure, which supports our commitment to sustainability while also delivering the performance and scalability required for the next generation of AI workloads.”
