KAIST's internal chip cooling tech slashes energy use by 90% for AI data centers

KAIST Develops Internal Cooling Solution for High-Performance Chips

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a liquid cooling technology that addresses one of the most pressing challenges facing AI infrastructure today: excessive heat generation in high-performance computing systems 1

1

. The breakthrough centers on embedding microscopic liquid-cooling channels, thinner than a human hair, directly inside silicon semiconductor chips rather than relying on external cooling mechanisms 2

2

. This approach to semiconductor cooling fundamentally changes how the industry might tackle thermal constraints in future data centers.

The rapid advancement of AI chips has created an energy crisis for data centers, which consume massive amounts of electricity both for intensive computations and for cooling the resulting heat. Traditional cooling methods such as air fans and external copper heat spreaders are reaching their physical limits, pushing the industry to seek alternatives 1

1

. Professor Sung Jin Kim noted that as the performance of AI semiconductors becomes increasingly limited by heat, this technology could serve as a foundational cooling solution for future high-performance computing 1

1

.

How Manifold Microchannel Structure Outperforms External Systems

Most modern liquid cooling systems rely on an external cold plate pressed against the outside of a chip, forcing coolant to travel from one end of the hardware to the other. This long journey creates immense fluid resistance, requiring heavy-duty pumps pushing with massive pressure that consumes substantial energy 1

1

. The KAIST team's manifold microchannel structure takes a different approach by featuring multiple tiny inlets and outlets scattered uniformly across the chip, mimicking an efficient logistics network 1

1

.

This decentralized design shortens the fluid's travel distance, reducing flow resistance and pumping pressure. The researchers used a multi-fidelity optimization framework to perfectly balance channel dimensions and flow rates, combining rapid 1D computational models with heavy-duty simulations to map optimal, uniform flow 1

1

. Previous manifold microchannel cooling systems often suffered from uneven coolant distribution, but the KAIST team optimized the structure so coolant moves more evenly throughout the system 2

2

.

Record-Breaking Coefficient of Performance for AI Data Center Cooling

The experimental results exceeded expectations dramatically. The system registered a Coefficient of Performance of 106,000, meaning a single unit of energy used for cooling can remove 106,000 units of heat 2

2

. This metric is ten times higher than the previous world record published in Nature in 2020 1

1

. In practical terms, chip manufacturers need just one-tenth of the pumping power to remove the same amount of heat, potentially transforming AI data center cooling economics.

Even under extreme thermal loads of 2,000 watts per square centimeter, the system kept chips comfortably below 100°C 1

1

. Tests applying the design principle to data center cold plates showed cooling performance improvements of more than 30 percent compared with conventional systems 2

2

.

Internal Cooling with Room-Temperature Water Changes Manufacturing Economics

Perhaps the most disruptive aspect of this development is its simplicity and compatibility with existing infrastructure. The system operates using internal cooling with room-temperature water and does not require boiling-based cooling methods, nanostructured surfaces, or expensive materials such as diamond 2

2

. The entire fabrication process happens below 350°C, making it entirely compatible with existing commercial semiconductor manufacturing lines 1

1

.

Foundries can integrate this technique for cooling high-performance chips into current designs without investing billions of dollars in new machinery 1

1

. The technology can be integrated into existing semiconductor manufacturing processes without additional production facilities, according to the research team 2

2

. The findings were published in the journal Energy Conversion and Management 1

1

2

2

.