NVIDIA completely abandons air cooling in AI servers

NVIDIA, one of the world's largest technology giants, has implemented a revolutionary change in its AI infrastructure based on the new Rubin architecture. Powerful computing systems will now be cooled entirely by liquid instead of traditional air cooling. This technology not only increases server efficiency but also significantly reduces the energy consumption of data centers. This is reported by Ixbt.com reports .

According to ixbt.com, the heat-transfer fluid in the new system operates at a temperature of 45 °C. This is a radical departure from the traditional approach, as previously, cooling servers required large volumes of cold air, powerful fans, and energy-intensive refrigeration units. In the Rubin architecture, the liquid circulates through a closed loop, removing heat directly from the components themselves.

A new milestone in energy and water efficiency

Experts estimate that increasing the cooling system temperature by just 1 °C can reduce energy costs by approximately 4 percent. For example, for a large data center with 50 MW of power, this means savings of nearly 4 million dollars per year. The new solution presented by NVIDIA is aimed at exactly this level of economic efficiency, significantly lowering the maintenance costs of AI systems.

Furthermore, the new architecture minimizes water consumption. Systems with traditional cooling towers can consume 2.6 million gallons of water per year for every 1 MW of power. NVIDIA's closed-loop liquid cooling system allows this figure to be reduced to almost zero. This is particularly important in the current era, where environmental sustainability and resource conservation are critical issues.

Compactness and silent operation

The primary goal of the system is not to keep chips ice-cold, but to ensure they operate stably within a safe temperature range. The liquid enters the processor at 45 °C and rises to 55 °C after absorbing heat. Special cooling plates keep components within safe limits, while external dry coolers dissipate heat without the need for constant operation of mechanical cooling systems.

In addition to energy and water savings, the Rubin architecture makes servers much more compact and silent. A system that previously occupied 6 rack units (U) now fits into just 2 units. This allows for efficient use of space in data centers and denser infrastructure deployment. NVIDIA plans to build all future AI centers based on this highly efficient architecture.