NVIDIA has introduced a new line of AI servers that utilize water cooling at temperatures up to 45 degrees Celsius (113 degrees Fahrenheit). This method contrasts sharply with traditional cooling systems, which rely heavily on energy-intensive air chilling and mechanical chillers.
The new servers, part of NVIDIA’s Rubin architecture, offer a fully liquid-cooled solution, eliminating the need for fans, cold aisles, and the extensive air cooling systems commonly used in data centers. Cooling has historically consumed about 40% of a data center’s electricity bill, and adjusting chiller temperatures can reduce cooling costs. For instance, raising chiller temperatures by just one degree can result in a 4% decrease in energy expenses, translating to savings of approximately $4 million annually for a 50-megawatt facility.
NVIDIA’s system significantly reduces water usage, previously a concern for data centers. Typical cooling towers can consume about 2.6 million gallons of water per megawatt per year. In contrast, NVIDIA’s new liquid cooling architecture brings this figure down to nearly zero, effectively eliminating the need for evaporative cooling, water towers, and continuous water drainage.
While the data center industry has long believed that lower temperatures equate to safer operations, NVIDIA’s data suggests otherwise. The coolant enters at 45 degrees Celsius and exits at 55 degrees Celsius, maintaining safe operating conditions. However, the effectiveness of this cooling method does depend on geographic factors, as temperature management can vary significantly in different climates.
The Rubin architecture also addresses noise issues typical of data centers, as it operates quietly without the need for fans and other noisy equipment. Additionally, this innovative server design is more space-efficient, with systems that previously occupied six rack units now fitting into just two.
NVIDIA’s advancements aim to mitigate several challenges inherent to data center operations, including excessive energy and water consumption, noise pollution, and spatial inefficiencies. While these developments do not rectify the existing infrastructure, they are anticipated to inform the design of future data centers.
