Nvidia has introduced a warm-water cooling design that could eliminate nearly all water consumed for cooling inside some AI data centres. However, electricity generation and chip manufacturing may continue to give these facilities a significant water footprint beyond their walls.
The company says its latest AI systems can operate with cooling liquid entering server racks at 45°C, or 113°F. This allows operators in suitable climates to release heat into the outside air without relying on water-consuming evaporative cooling towers.
Closed-Loop System Reduces On-Site Consumption
The coolant passes directly through the servers, absorbs heat from the chips, and exits at approximately 55°C. It then releases that heat through external radiators before returning to the racks.
Because the liquid circulates through a closed loop, the system generally needs to be filled only once. Nvidia says this could reduce on-site water consumption for cooling by as much as 100% under favourable operating conditions.
The higher coolant temperature could also reduce or eliminate the need for mechanical chillers and, in some facilities, fans. That would lower electricity use, reduce noise, and simplify cooling infrastructure.
Nvidia Chief Sustainability Officer Josh Parker said the technology had largely addressed the water-consumption challenge facing data centres. The company’s claim, however, focuses mainly on water used within the facility for cooling its computing equipment.
Power Generation Adds Indirect Water Demand
Data centres also consume water indirectly through the electricity used to operate servers and other equipment. Thermal power stations commonly use water to cool machinery and release excess heat.
The US Geological Survey estimated that thermoelectric power plants consumed approximately 2.7 billion gallons of water per day in 2015. The amount varies substantially depending on the fuel, cooling design, location, and operating conditions.
Natural gas and coal plants generally use considerably more water than wind and solar generation. Hydroelectric facilities can also have a substantial water footprint because water evaporates from reservoirs, although estimates depend on how those losses are allocated.
The International Energy Agency expects natural gas and coal to meet more than 40% of the additional electricity required by data centres through 2030. Global data-centre power consumption is projected to reach approximately 945 terawatt-hours by that year.
Cooling Solves Only Part of the Problem
Water is also used in semiconductor manufacturing, where producers require large quantities of highly purified water to clean wafers during production. Nvidia’s cooling system does not reduce that upstream demand.
The technology could therefore deliver a major improvement at the facility level without eliminating the full water footprint of AI computing. Its total impact will depend partly on whether data centres receive their electricity from water-intensive thermal plants or lower-water sources such as wind and solar.
Featured image credits: Roboflow Universe
For more stories like it, click the +Follow button at the top of this page to follow us.