Cooling data centres: How pumps and valves enable artificial intelligence

AI will drastically drive up power consumption

Artificial intelligence is a power muncher: Compared to a search in Google, consuming 0.0003 kilowatt hours (kWh) of power, a ChatGPT-3 prompt is estimated to use 0.0029 kWh – almost ten times as much. The International Energy Agency (IEA) predicts that the global annual power consumption of data centres, cryptocurrencies and artificial intelligence will have risen from 460 terawatt hours (TWh) to anywhere between 620 and 1050 TWh by 2026. That equals about two percent of the world's energy consumption, an amount comparable with the power consumption of a country of a size between Sweden and Germany.

The development of the energy consumption of AI servers illustrates the speed at which power consumption is growing because of artificial intelligence. The DGX H100 system by NVIDIA is currently considered one of the most powerful servers for AI tasks. It uses up to 10.2 kilowatts, which is about as much as a professional electric stove in a commercial kitchen. Its predecessor, DGX A100, which was taken off the market in January 2024, only used a maximum of 6.5 kilowatts. The latest system, DGX B200, which NVIDIA announced in March 2024, will use as much as 14.3 kilowatts.

In some countries, the energy requirements of data centres will take on extreme dimensions: In Ireland, for example, the share of power consumed by data centres already equals 17 percent of the country’s overall power consumption. According to an IEA estimate, this will rise to 32 percent by 2026 – bringing with it the corresponding challenges regarding the country’s climate targets and its infrastructure. This is where pumps and valves can contribute to a solution.

Water is the more efficient coolant

AI applications will concentrate more and more power in individual server racks, generating progressively more heat and requiring increasingly efficient cooling. This is where air as a coolant hits its limits; new types of temperature control will become necessary. One option of improving cooling is to increasingly use water instead of air. Compared to air, water can absorb four times the amount of energy.

The more a data centre uses a liquid for cooling and the closer this liquid is to the processors, the lower the energy input. With artificial intelligence on the rise, the focus has shifted to cooling methods whose heat exchangers are located directly in the server racks (rear door cooling or in-rack cooling) or processors that are fitted with their own cooling circuits – similar to gaming PCs (direct liquid cooling or direct-to-chip cooling). The US company Tesla, for instance, has just started an AI data centre in Texas that uses 50,000 NVIDIA chips. When fully developed, the centre’s power consumption is predicted to reach 500 megawatts. They opted for direct-to-chip cooling. According to the manufacturer, this is going to reduce the power costs for cooling by 89 percent.

Holistic system optimisation saves the most energy

Data centre operators have to augment the efficiency of cooling to master the challenges of artificial intelligence. This is achieved by opting more for liquid cooling and making these cooling systems as efficient as possible. High efficiency of cooling circuits goes beyond the high efficiency of its individual components such as heat exchangers, pumps and valves. The components also have to be matched to each other to make sure they work together efficiently as a system. Precisely this holistic view of systems is KSB’s strength.

Figures demonstrate just how important optimising the overall system is. Ten percent of the energy consumption of a cooling circuit can be saved by installing high-efficiency pumps, such as KSB’s Etanorm, for example. KSB’s engineers have optimised this pump’s impeller to provide an optimum efficiency and prevent a drop in performance through cavitation also at low pressure. Its synchronous reluctance motor meets the requirements of the highest energy efficiency class IE5 (IEC/TS 60034-30-2); compared with IE4, motor losses are reduced by a further 20 percent.

Another ten percent of energy can be saved by using a frequency inverter for controlling the pump motors. KSB’s PumpDrive, for example, estimates the current operating point on the basis of the motor input power and speed. It identifies areas of motor inefficiency, such as extreme part load, dry running or overload, and adjusts the speed to the actual requirements.

However, by far the greatest potential for optimisation lies in improving the entire system – up to 60 percent in total. This percentage of energy can be saved by determining the actually required system pressure, for example, in order to dimension the pumps and impellers accordingly or prevent pressure losses by valves or bottlenecks. 

KSB’s expertise ensures an optimum selection of cooling systems. An example is the FluidFuture® service as part of which KSB experts make use of potential savings by conducting measurements, selecting suitable high-efficiency pumps and components and taking care of proper installation and commissioning. The FluidFuture® energy saving concept also comprises the operation of the system with continuous monitoring.