NVIDIA Cools Its AI Servers with Liquid at 45°C: A Leap in Efficiency

What Happened?

NVIDIA has revealed that its Rubin generation AI servers can operate with liquid cooling at temperatures up to 45°C (113°F), exceeding the typical range of a hot tub (38-40°C). This is the world's first system to achieve 100% liquid cooling, without fans, covering every chip and network component. The design is detailed in the NVIDIA DSX reference, a guide for building complete AI factories. According to NVIDIA's blog, this closed-loop system eliminates the need for evaporative cooling, reducing water consumption to nearly zero. The Rubin generation offers a generational leap in efficiency, as each degree increase in coolant temperature reduces cooling costs by 4%, according to industry estimates.

Why Is This Important?

Historically, cooling accounts for up to 40% of a data center's electricity consumption, according to a 2022 McKinsey report. By raising the liquid coolant temperature to 45°C, NVIDIA enables the use of dry coolers for most of the year, eliminating the need for chillers and reducing water consumption to nearly zero. According to NVIDIA's blog, a 50 MW data center can save over $4 million annually in energy and water costs. Additionally, the water footprint drops from 2.6 million gallons per MW per year to virtually zero. This advance is crucial in a context where data centers consume about 1% of global electricity and their water demand competes with local communities in arid regions. Ali Heydari, director of data center cooling at NVIDIA, stated: “With dry cooler-based designs, it's a closed-loop system without evaporative water cooling, except perhaps 1% of the year when we might need chillers in some climates.”

Consequences for the Industry

This breakthrough accelerates the transition toward more sustainable data centers. The Rubin generation offers more compute power per watt, and by eliminating fans and reducing reliance on traditional cooling systems, operational complexity decreases. Companies like Microsoft, Google, and Amazon, which are already investing in liquid cooling, could adopt this architecture for their AI clusters. However, implementation requires redesigning facilities, which poses an initial barrier. In the long term, reduced operating costs and compliance with environmental regulations will make this technology indispensable. According to a 2023 Uptime Institute report, 40% of data center operators already use or plan to use liquid cooling within the next two years. NVIDIA's solution could accelerate that adoption, especially in water-restricted regions like the southwestern United States or parts of Europe.

What Readers Should Know

• Energy efficiency: Each degree increase in coolant temperature reduces cooling costs by 4%, according to industry data cited by NVIDIA.
• Zero water consumption: The system is a closed loop with dry coolers, no evaporation, except in extreme climates where chillers may be required less than 1% of the year.
• No fans: All components are liquid-cooled, eliminating fan noise and power consumption, which typically accounts for 10% to 15% of the cooling system's energy use.
• Scalability: The DSX design is intended for complete AI factories, from individual racks to hyperscale installations. NVIDIA states the guide covers everything from planning to operation, including integration with renewable energy sources.

“Liquid cooling at 45°C not only reduces energy consumption but virtually eliminates water usage, a resource increasingly scarce in many regions.” — Ali Heydari, director of data center cooling at NVIDIA.

Context and Comparisons

Companies like Microsoft have experimented with liquid cooling by submerging servers in dielectric fluids, while Google uses recycled water in some data centers, such as its Hamina, Finland facility, which uses cooled seawater. However, NVIDIA's solution stands out for its simplicity and efficiency by operating at higher temperatures. Compared to traditional air cooling, which requires large volumes of air conditioning and consumes up to 30% of a data center's energy, liquid cooling at 45°C allows heat to be dissipated directly outdoors via dry coolers, even in warm climates. For example, in a climate like Phoenix, Arizona, where summer temperatures exceed 40°C, dry coolers can keep the liquid temperature below 45°C most of the year, according to NVIDIA simulations. This contrasts with air cooling systems that require compressors and refrigerants with high global warming potential.

Impact on the Future of Work and AI

By reducing data center operating costs, AI becomes more accessible and sustainable. Companies will be able to train larger models with a lower carbon footprint. Additionally, eliminating fans reduces noise, allowing data centers to be located in urban areas with noise restrictions. This could accelerate AI adoption in sectors like healthcare, logistics, and manufacturing. According to a University of California study, training a large language model like GPT-3 emits approximately 552 tons of CO2 equivalent. With NVIDIA's cooling, that footprint could be reduced by 30% from energy efficiency alone. Moreover, water savings are critical: a typical 50 MW data center consumes about 130 million gallons of water per year, equivalent to the consumption of 1,200 U.S. households. Eliminating that consumption alleviates pressure on local water resources. In the future of work, the ability to deploy AI clusters in urban locations could democratize access to high-performance computing, allowing startups and SMEs to compete with tech giants.