KAIST Reduces AI Data Center Cooling Power by 90% [Reading Science]

Implementation of Ultra-Fine Microfluidic Channels Inside Semiconductor Chips Achieves Cooling Efficiency Ten Times Higher Than the Global Standard
"In the AI Era, the Key Battleground Is Heat Management"—Potential for Application to NVIDIA's Next-Generation Chips

A breakthrough ultra-high-efficiency semiconductor cooling technology capable of drastically reducing the power consumption of AI data centers has been developed by a Korean research team. By implementing microfluidic channels thinner than a strand of hair inside semiconductor chips, the team achieved a cooling efficiency ten times higher than the previous world-leading standard, presenting a new solution to the heat dissipation issue widely regarded as the greatest challenge for AI semiconductors.

KAIST announced on June 16 that a joint research team led by Professor Sungjin Kim from the Department of Mechanical Engineering and Professor Ikjin Lee from the Department of AX has developed an ultra-high-efficiency liquid cooling technology that combines a manifold with microchannels inside semiconductor chips.

Structure of manifold microchannel cooling device for high-heat semiconductor chip cooling. Provided by research team

As the proliferation of generative AI intensifies competition for AI semiconductor performance, heat management has emerged as the industry's top issue. In particular, AI data centers are known as "power hogs" due to the massive amount of electricity consumed for cooling. To overcome the limitations of air-cooling methods, the industry is accelerating efforts to introduce liquid cooling technologies.

The technology developed by the research team is an improved version of the manifold microchannel (MMC) structure, in which coolant is circulated directly inside the semiconductor chip to remove heat. The manifold distributes and collects coolant through multiple pathways, while the microchannels are ultra-fine fluid pathways thinner than a human hair.

To address the issue of coolant concentrating in only some channels in conventional MMCs, the team optimized the structure so that coolant flows evenly through all channels. Similar to how parcel delivery networks shorten delivery distances by distributing packages to multiple regional logistics centers, the optimized design minimizes coolant travel distance, reducing energy loss.

In the AI Era, Heat Management Determines Competitiveness—Not Just Computing Power

Experimental results using a silicon wafer showed a coefficient of performance (COP) of 106,000. This means that for every unit of energy spent on cooling, 106,000 times that amount of heat can be removed.

This value is more than ten times higher than the previous world record reported in the international journal "Nature" in 2020. In other words, the amount of energy required to remove a given quantity of heat can be reduced to one-tenth of the previous level.

A key advantage of this technology is that it was achieved using only room-temperature water—without the need for expensive materials such as diamond, boiling-based cooling, or nano-structured surface treatments. It is also compatible with current semiconductor manufacturing processes, making it highly feasible to implement without additional investment in equipment.

Research team photo. (Top row) Corresponding author Professor Sungjin Kim, Department of Mechanical Engineering, Professor Ikjin Lee. (Bottom row) First author Dr. Youngjin Lee, PhD candidate Hansol Lee, PhD candidate Chulhyun Hwang. Provided by KAIST

After verifying the technology on a test chip measuring 5mm by 5mm, the research team explained that it can be applied to large semiconductors for AI data centers. When applied to an actual data center cooling device called a cold plate, it demonstrated a cooling performance improvement of more than 30% compared to existing solutions.

The team expects that the technology could be applied to next-generation ultra-high-performance chips, such as NVIDIA's upcoming AI platform, Vera Rubin.

Sungjin Kim, Professor at KAIST, commented, "The latest AI semiconductors generate hundreds or even thousands of watts of heat from a single chip, making air-cooling alone insufficient. Going forward, the competitiveness of AI data centers will depend not only on computing power, but also on how efficiently they manage heat."

The results of this research were published on June 15 in the international journal "Energy Conversion and Management." The study was supported by the Mid-Career Researcher Program of the National Research Foundation of Korea and by the Defense Technology Promotion Research Institute’s Specialized Laboratory for Ultra-High Heat Flux Cooling Systems.

This content was produced with the assistance of AI translation services.