As power consumption at artificial intelligence (AI) data centers increases, researchers have developed a technology that can significantly reduce the power needed to cool semiconductors.
KAIST said on the 16th that a research team led by Mechanical Engineering Professor Kim Seong-jin and AX Department Professor Lee Ik-jin developed a liquid cooling technology that creates micro water channels thinner than a human hair inside semiconductor chips to remove heat directly. The research was published on the 15th (local time) in the international journal Energy Conversion and Management.
The team improved conventional manifold microchannel cooling technology to design coolant flow evenly through multiple passages inside the chip. Microchannels are very fine water channels created inside the chip, and the manifold is a structure that splits the coolant into multiple branches and then recombines it. Supplying coolant at multiple points instead of sending it in a long flow from one side shortens the travel distance and reduces the energy required for cooling.
Existing technology had the problem of coolant concentrating in some passages. The team used computational models and precision simulations to find a structure in which coolant flows uniformly through all channels, and they fabricated it on an actual silicon wafer to verify performance.
In experiments, the coefficient of performance (COP), which indicates cooling efficiency, reached 106,000. COP shows how much heat can be removed compared with the energy input, and 106,000 means that with one unit of energy used for cooling, an amount of heat equivalent to 106,000 units can be removed. The team said this figure is more than 10 times higher than the previous top level reported in Nature in 2020.
In particular, this technology was implemented using only water at room temperature, without complex methods that rely on boiling phenomena, nano-surface treatment, or expensive materials such as diamond.
The results were verified on a 5 mm × 5 mm test chip. The team believes the same principle can be applied to large AI Semiconductor chips such as graphics processing units (GPU) and tensor processing units (TPU). Applied to a data center cold plate, it showed more than a 30% improvement in cooling performance over existing methods.
Kim Seong-jin said, "In the AI era, not only semiconductor performance but also how effectively we control heat is important," adding, "We hope this technology will be used to reduce power consumption at data centers."
References
Energy Conversion and Management (2026), DOI: https://doi.org/10.1016/j.enconman.2026.121422