
Steam cooling embedded in hot chips
Researchers in Japan have shown how microfluidic channels embedded in AI processors and GPUs can provide significant cooling using a dual phase approach.
The team at the Institute of Industrial Science at the University of Tokyo used microchannels embedded directly into the chip itself. These channels allow water to flow through, efficiently absorbing and transferring heat away from the source with dual phase cooling.
Current microfluidic techniques are limited by the heat capacity of water. Using the latent heat of phase change of water, which is the thermal energy absorbed during boiling or evaporation, provides seven times the amount of cooling.
“By exploiting the latent heat of water, two-phase cooling can be achieved, resulting in a significant efficiency enhancement in terms of heat dissipation,” said researcher Hongyuan Shi.
However this is a challenge to manage the flow of vapour bubbles as the water turns to steam. Maximizing the efficiency of the heat transfer depends on a variety of factors, including the geometry of the microchannels, the two-phase flow regulation, and the flow resistance.
The water-cooling system developed by the team uses three-dimensional microfluidic channel structures with a capillary structure and a manifold distribution layer. The researchers designed and fabricated various capillary geometries and studied their properties across a range of conditions.
The manifolds incorporate a two-stage flow design. The coolant enters the spacious manifold channels, then is forced to flow through narrow microchannels in contact with the heated surface, and exits through another manifold channel. The structure of the manifolds controls the coolant distribution in the microchannel.
The team compared the flow boiling between a microchannel with a diameter of ∼33 μm without a manifold and four samples with identical microchannels and different manifolds (hydraulic diameter of ∼240 μm) to assess the cooling performance. The microchannels have a length of 2,700 μm, a depth of 100 μm, and a width of 20 μm.
They found that both the geometry of the microchannel and the manifold channels which control the distribution of coolant influence the thermal and hydraulic performance of the system.
The ratio of useful cooling output to the required energy input, known as the coefficient of performance (COP), reached up to 100,000, significantly greater than conventional cooling techniques.
“Thermal management of high-power electronic devices is crucial for the development of next-generation technology, and our design may open new avenues for achieving the cooling required,” says Masahiro Nomura at the Institute.
The microchannels could be etched in silicon and wafer bonded to the back of a chip with the microchannels designed to focus on the hotspot areas in AI chips to provide more effective cooling in AI datacentres.
The paper is at 10.1016/j.xcrp.2025.102520
