Advances and Challenges in Liquid Cooling for Data Centres Amid Rising AI Demand
Iceotope has developed a liquid cooling system that uses a closed-loop water circuit to cool an oil-based fluid and other components, claiming up to an 80% reduction in energy demands related to cooling. Some data-centre operators continuously overclock servers using liquid cooling, with a US hotel chain planning to repurpose server heat for heating guest rooms, laundry, and even a swimming pool.
A recent incident highlights the importance of reliable cooling: a November failure of the cooling system at CME Group disrupted trading technology, prompting the group to add external cooling capacity to reduce the risk of recurrence. Demand for data centres powered by AI is rising amid environmental concerns. More than 200 groups have called for a moratorium on new data centres, while companies seek ways to minimize power and water usage while pursuing silent operation.
Certain two-phase cooling systems use refrigerants containing PFAS, which pose safety and climate risks. Some designs now incorporate PFAS-free refrigerants as safer replacements, addressing issues such as vapor leakage. Microsoft's subsea data centre near Orkney demonstrated impressive performance with a PUE of about 1.07 and zero water usage. However, the project was shut down due to unfavourable economics. Insights from this project are informing ongoing research into liquid cooling technologies.
Researchers at the University of California San Diego are exploring a pore-filled membrane cooling technology that could passively cool chips without active pumping, showing potential for commercialisation. Experts highlight that AI models consume substantial energy and call for increased transparency about their energy usage. Industry leaders, including Microsoft, are discussing the limits and productivity of large language models amidst these concerns.
