Innovations and Challenges in Data Centre Cooling Amid Rising AI Demand
Data centres are increasingly adopting liquid cooling and two-phase cooling techniques, including showering and bathing components, to manage heat and enable high-speed, overclocked operation.
Iceotope, a company in this field, claims its liquid cooling technology can reduce cooling-related energy demand by up to 80%. Their system uses a closed water loop to avoid drawing additional water from local supplies, and employs partly fossil-fuel-derived but PFAS-free oil-based fluids.
A US hotel chain is innovating by planning to use heat generated from hotel data-centre servers to warm guest rooms, laundry facilities, and a swimming pool, demonstrating creative reuse of waste heat.
However, cooling-system failures can have significant impacts, as seen with CME Group in November when a failure disrupted trading technology. To prevent recurrence, CME added external cooling capacity.
The rising demand for data centres, fueled by growing AI workloads, is controversial due to their high energy and water consumption. Over 200 environmental groups in the US have called for a moratorium on new data centre developments.
Two-phase cooling systems sometimes rely on PFAS-containing refrigerants, which pose safety and greenhouse-gas concerns. While some firms are transitioning to PFAS-free refrigerants, others continue their use.
Microsoft's subsea data centre off Orkney utilized cold seawater for cooling, achieving a low power usage effectiveness (PUE) of 1.07 with zero water consumption. Despite this, the project was discontinued due to unfavorable economics. Nevertheless, Microsoft is exploring advanced liquid cooling techniques such as microfluidics.
Researchers at UC San Diego, including Chen, proposed a novel pore-filled membrane cooling technology that allows passive cooling by eliminating the need for active pumping; this technology has potential for commercialization.
Given that AI models, especially large language models (LLMs), are highly energy-intensive and increase cooling demands, industry experts call for greater transparency from AI companies regarding their energy usage. Some authorities suggest that LLMs may have reached their productivity limits, underscoring the need for efficiency improvements.
