NVIDIA Cools Its Chips with Liquid at 45°C: Lower Consumption in Data Centers

What Happened?

NVIDIA has revealed that its next-generation AI infrastructure, called Rubin, will incorporate a liquid cooling system that operates at a temperature of 45°C. This value exceeds the typical temperature of a jacuzzi (around 40°C) and represents a radical shift from traditional air or cold liquid cooling approaches. The company claims that Rubin will be the first platform to achieve 100% liquid cooling, completely eliminating fans and using a closed loop that captures heat directly on the surface of components. According to NVIDIA's official blog, the system covers not only accelerators and CPUs but all chips and network components within a sealed circuit that prevents water evaporation. The announcement was made at the GTC 2024 conference, where Jensen Huang detailed the company's roadmap through 2026.

Why Is It Important?

NVIDIA's decision addresses a critical challenge: the energy consumption of AI data centers has skyrocketed due to increased computational density. According to the company, efficiency is measured not just by energy consumed, but by the computational work completed per kilowatt-hour. By operating the coolant at 45°C, the system can dissipate heat using dry coolers (large radiators) instead of relying on mechanical chillers, significantly reducing the electricity dedicated to cooling. Additionally, being a closed loop, water does not evaporate, reducing water usage in facilities. This is especially relevant in water-stressed regions like the southwestern United States or parts of Chile. The International Energy Agency (IEA) estimates that data centers consumed between 1% and 1.5% of global electricity in 2023, and that figure is expected to double by 2026 due to AI. NVIDIA claims its approach can reduce cooling energy consumption by up to 30%, based on internal models and tests with customers like CoreWeave.

Consequences for the Industry

This innovation could redefine data center design. By eliminating fans and reducing the need for HVAC systems, operators could save up to 30% in energy costs, according to analyst estimates. However, implementation requires changes to existing infrastructure, such as installing pipes and liquid distribution units. Major tech companies like Google, Microsoft, and Amazon are already experimenting with liquid cooling, but NVIDIA's approach of using hot liquid could accelerate mass adoption. Google, for example, has used liquid cooling in its TPUs since 2018, but at lower temperatures (around 20°C). Microsoft has tested immersion in dielectric liquid for its Azure servers, but not at 45°C. NVIDIA's advantage lies in its solution being integrated from chip design, optimizing thermal transfer. Data center companies like Equinix and Digital Realty are already evaluating how to adapt their facilities for Rubin, according to industry sources cited by The Register. However, the initial cost of retrofitting facilities could be a barrier for smaller operators, as installing pipes and pumps can add 10% to 20% to the construction cost of a data center, according to Uptime Institute estimates.

What Readers Should Know

• Not speculation: NVIDIA has presented Rubin as a real product, with availability expected in 2026. The company has already shown prototypes to select partners, according to AnandTech sources.
• Environmental impact: Reducing water consumption is key in water-stressed regions. A typical data center can consume between 3 and 5 million liters of water per year for cooling; NVIDIA's closed system almost completely eliminates that demand.
• Technical challenges: Liquid at 45°C requires special materials and seals to prevent leaks and corrosion. NVIDIA uses a mixture of deionized water and anticorrosive additives, similar to that used in high-performance automotive cooling systems, but with tighter tolerances.
• Competition: AMD and Intel are also developing liquid cooling solutions, but none have announced such high temperatures. AMD, for example, has presented its MI300X platform with optional liquid cooling, but at standard temperatures of 30-35°C. Intel is working on two-phase immersion systems, but still in experimental phase.

Context and Comparisons

Historically, data centers have been cooled with air conditioning or cold water (around 10-20°C). The 45°C temperature represents a paradigm shift, similar to when Google began using canal water to cool its servers in 2009, reducing PUE (Power Usage Effectiveness) from 1.5 to 1.1. However, NVIDIA's approach is more ambitious, aiming to completely eliminate traditional cooling systems. Another comparable milestone was the introduction of immersion cooling in 2017 by 3M and Green Revolution Cooling, which submerged servers in dielectric liquid, but at lower operating temperatures (around 25°C). NVIDIA's innovation is based on increasing the coolant temperature, allowing the use of dry coolers instead of mechanical chillers, saving energy. According to a University of Texas study, each degree increase in coolant temperature can reduce cooling energy consumption by 3-5%. Thus, moving from 20°C to 45°C could represent a 75-125% savings, although NVIDIA has not provided exact figures.

"Energy efficiency is not just about consuming less energy, but about completing more computational work per kilowatt-hour used" — NVIDIA

This quote reflects the philosophy behind the innovation: optimizing performance per watt, not just reducing absolute consumption. In practical terms, Rubin could achieve an efficiency of 10 TFLOPS per watt (in FP16 precision), surpassing the 8 TFLOPS/watt of the previous Hopper generation, according to leaked specifications.

Implications for the Future

If NVIDIA's technology proves viable, it could accelerate the construction of denser and more powerful data centers needed for the next generation of AI models. For example, training models like GPT-4 required around 50 GWh of electricity; with Rubin, that consumption could be reduced by 30% thanks to improved cooling. Additionally, reducing water and electricity consumption aligns the industry with global sustainability goals, such as the EU's commitment to carbon neutrality by 2050. However, the initial cost of retrofitting facilities could be a barrier for smaller operators. There is also the risk that the 45°C temperature could shorten the lifespan of electronic components, although NVIDIA assures that stress tests have shown reliability comparable to air-cooled systems. The company is also developing an intelligent monitoring system that adjusts liquid flow in real time based on workload, minimizing the risk of hot spots. In summary, Rubin represents a bold step toward more efficient and sustainable data centers, but its success will depend on industry adoption and NVIDIA's ability to resolve associated technical challenges.