Schneider says hotter coolant can make AI datacenters less thirsty

Schneider Electric's white paper, "Water Usage at AI Scale: Insights from a 100 MW Comparative Analysis," asserts that the cooling architecture of datacenters significantly impacts their water and energy consumption. Comparing four 100 MW datacenter designs, the report concludes that transitioning from air cooling to liquid cooling consistently enhances energy efficiency and can decrease water consumption in locations where water is limited.

Schneider, having recently acquired Motivair, a liquid-cooling specialist, has a vested interest in these findings.

The paper models the performance of the same facility in Paris, France, and Dallas, Texas, revealing that the air-cooled design in Dallas has the highest onsite water usage. When liquid cooling is implemented with coolant at 45°C (113°F), water usage drops by at least 50%, according to the study. This suggests that location, operating temperature, and cooling architecture are crucial factors in determining a facility's water requirements, even as rack densities and the demands of AI systems intensify.

The report highlights that water consumption is primarily influenced by the equipment used to dissipate heat outside the data center. For a facility of similar size, cooling towers can use five to 20 times more water than dry coolers, the report states. The first design scenario employs air cooling, while the second utilizes liquid cooling at 32°C (90°F), a typical supply temperature for early AI datacenters.

The third scenario maintains the same cooling equipment but sets the supply temperature to 45°C (113°F). The fourth design reduces equipment count by optimizing the system for 45°C operation, thereby cutting capital expenditure. Higher coolant temperatures yield significant water savings by extending the range of conditions where cooler outside air can reject heat without mechanical cooling.

Schneider advises that water consumption should be evaluated early in the design process and consider alternative cooling sources to lessen demand on local water supplies. This issue is particularly contentious in the US, where datacenter water consumption has tripled in the past decade, leading to increased local opposition. In the UK, the government has faced criticism for overlooking water demand while promoting the construction of additional AI server farms.

Schneider emphasizes that water and energy efficiency should be considered together, as the optimal design depends on the relative importance of power usage effectiveness (PUE) and water consumption, as well as local energy and water costs.