The rapid growth of modern data centres has significantly increased heat generation from computing equipment, placing greater demands on cooling systems. Traditional air-based cooling methods, which have long been used to regulate data centre temperatures, are becoming less effective as computing power and server density continue to rise. The increasing thermal output of high-performance systems, typically reaching heat loads of approximately 3 kW per unit, challenges the ability of conventional cooling approaches to maintain optimal operating conditions.
This rise in heat generation is further driven by the growing adoption of artificial intelligence technologies, which require higher processing capacity and result in increased power consumption and thermal output. As a result, there is a growing need for more efficient and reliable cooling solutions.
Today’s AI accelerators, graphics processors and dense multi-core CPUs can easily release 400–600 watts from a chip barely larger than a postage stamp. Traditional air cooling is hitting fundamental physical limits: air conducts heat very poorly and requires massive airflow which means noisy fans, high auxiliary electricity consumption, and rapidly worsening Power Usage Effectiveness (PUE).
Liquid cooling has emerged as a promising alternative to traditional cooling methods. By providing improved heat transfer capabilities and enhanced thermal management, liquid cooling systems offer a potential solution to address the limitations of conventional air-based systems and support the continued expansion of high-density data centre infrastructure. This article examines the thermal performance of traditional cooling methods and liquid cooling technologies in addressing the evolving cooling demands of modern data centres.
Cooling accounts for 30–45 % of electricity use in data centers. In South Africa, where power remains costly and supply is frequently constrained, that figure directly influences whether large-scale high-performance computing (HPC) can be affordably expanded.
South Africa is leading the way in computational research and has delivered one of the most practical results worldwide to stop supercomputers from overheating.
Detailed fluid dynamics simulations that were conducted to compare four different non-conductive dielectric liquids suitable for single-phase immersion cooling. This technique involves fully submerging heat-generating high-performance computing (HPC) components, such as CPU and GPU, in a cooling fluid to enable direct heat transfer. The results demonstrate that immersion cooling provides significantly higher cooling efficiency and improved thermal management compared to conventional air-based cooling methods.
The standout performer when both cooling ability and real-world cost were considered turned out to be ordinary electronic-grade mineral oil, the type already widely used in electrical transformers. Based on the pricing data from 2025, electronic-grade mineral oil—at approximately R123 per litre is 20–25 times more cost-effective than specialized products such as 3M Novec 7000 (≈R2,964/L) or 3M Fluorinert FC-72 (≈R2,470/L), while also demonstrating the highest heat absorption capacity of all fluids tested at 22.5 kJ/kg.
By transitioning to liquid immersion technologies like mineral oil-based systems, data centres can achieve substantial energy savings, quieter operation, reduced PUE, and greater scalability critical factors for sustaining AI-driven growth and HPC expansion in power-constrained regions like South Africa.
Article by: Sandile Khumalo-Mechanical Engineer