Google has unveiled Project Deschutes, a fifth-generation 400 V DC power supply and liquid cooling system for next-generation racks designed to support the rapid advancement of AI. The company has been using 48 V DC power for the past decade, but moving to the new standard will increase the maximum power per rack from 100 kW to 1 MW.
Individual racks with AI systems are expected to consume over 500 kW by 2030. For example, the upcoming NVIDIA Rubin Ultra NVL576 super accelerator, which will appear in 2027, will be “packed” into a new-generation Kyber rack and consume about 600 kW. Google, one must assume, will develop its own modification of this accelerator, adapted to its data centers, as it has already done for the GB200 NVL72.
Using 400V allows the use of the supply chain used by the electric vehicle industry, which helps reduce costs and improve quality. Google is working with Meta✴ and Microsoft on the Mt. Diablo project, which is developing common standards for electrical and mechanical interfaces. The first version of the specifications (v0.5) will be available for industry comment in May 2025.
Image source: Google
The power subsystem in Mt. Diablo is located in a separate module (sidecar). This increases the useful space in the server racks, allowing them to be entirely dedicated to accelerators, and increases overall energy efficiency by approximately 3%, which is very significant on a hyperscaler scale. In the future, we are considering switching to direct distribution of high-voltage direct current inside the data center for even greater efficiency and increased density.
Image source: Google
With the sharp increase in chip power consumption, the use of liquid cooling has become inevitable. Over the past seven years, Google has deployed liquid cooling in more than 2,000 TPU Pod clusters. Liquid cooling was first used for TPU v3 AI accelerators, which appeared in 2018. The company uses water blocks, which allows almost doubling the density of computing power compared to air cooling. When moving from TPU v2 to TPU v3, this also allowed the cluster size to be quadrupled. Liquid cooling is also used for Ironwood accelerators (TPU v7).
Project Deschutes’ CDU architecture, which uses redundant heat exchangers and pumps, provides 99.999% availability. Google plans to hand over the fifth generation of Project Deschutes to the Open Compute Project (OCP) in 2025. The publication of specifications, design data, and operational recommendations will accelerate the mass adoption of the liquid storage system in the industry. The company is confident that joint efforts will help the industry cope with future challenges in the AI industry and scale computing power further.