With the increase in electronic devices applying high-performance artificial intelligence (AI), the cooling solutions for AI semiconductors that handle vast amounts of information are becoming increasingly important. As semiconductors strive for higher performance and smaller sizes, the significance of heat management has also grown. If cooling is not carried out properly, it will inevitably lower the performance of the semiconductors. Conversely, if the cooling system is excellent, it can save energy and secure more space inside the product.
Efficient heat management is essential for data centers processing AI computations. According to the International Energy Agency (IEA), the power consumption of data centers worldwide has increased by about 12% annually since 2019, reaching 415 terawatt hours in 2024. It is expected to rise to 945 terawatt hours by 2030, exceeding Japan’s total annual power consumption.
MHS is a domestic corporation specializing in AI semiconductor cooling solutions. MHS collaborates with leading AI semiconductor chip designers in Korea, such as FuriosaAI and Rebellion. An interview was conducted with MHS CEO Lim Jong-soo at the MHS headquarters in the Seoul Digital National Industrial Complex on April 15. CEO Lim noted, “There are almost no electronic products that do not use a cooler,” adding that “a transition is taking place in the cooling technology field comparable to the shift from wired to wireless telephones.” The following is a Q&A.
What is the motivation for starting the business?
I served as the Head of Team responsible for developing heat spreaders at Celcia Technologies and handled technical services to optimize product designs using computational fluid dynamics at Siemens Technology Research Institute. After simulating the self-designed cooling technology, I concluded that the performance improvement effect compared to existing products was so good that I immediately thought I needed to start a business.
Why is cooling important in the AI semiconductor field?
It’s simple. AI semiconductors use high-bandwidth memory (HBM). HBM is memory made by stacking multiple DRAMs. To allow data to move quickly back and forth between the memory semiconductors and the graphics processing unit (GPU), HBM is installed right next to the GPU. Due to the stacking structure of DRAM, the complexity of the process is high, and since the DRAM is grouped together, heat generation is significant. In the past, the heat generated by semiconductors was only about 100-200W, but with the evolution to AI semiconductors and HBM, the heat generation has skyrocketed to the kilowatt range. When 1 kilowatt of heat is generated from a semiconductor that is only 20-30mm in size, it cannot be managed with existing technology. The heat output of a small electric heater used in an office is 1 kilowatt. Currently, the number of stacked DRAM layers is at about 12 to 16, but when it reaches 24 or 48 layers, managing heat will become even more challenging. Expanding the area makes heat management easier, but due to the nature of semiconductor products, they cannot be made significantly larger. This is why cooling technology is important.
How is MHS’s technology different?
Traditional semiconductor cooling technology mainly used air cooling. Due to the cooling problems associated with AI semiconductors, cooling technology is transitioning from air cooling to liquid cooling, which uses coolant to dissipate heat. Consider that when you operate a hairdryer, air passes over heating wires and immediately produces hot air, whereas water, even when placed on a high flame, takes 5-10 minutes to reach 100 degrees. Water can absorb a significant amount of heat. In other words, water can remove a large amount of heat more quickly.
MHS utilizes a proprietary micro-channel liquid cooling method (MACS) among liquid cooling techniques. As the performance of semiconductors increases, heat generation becomes inevitable, and there is a limit to the liquid cooling technology’s ability to enhance heat performance by increasing flow rates. The characteristic of MHS’s MACS is that it converts water flow pipes into several pipes with a diameter of less than 1mm. It is designed to quickly remove heat from the chip and maximize cooling performance, especially beneficial for heat sources like flat semiconductor chips with a small heat dissipation area. It took 10 years to develop and mass-produce this technology.
Immersion cooling is emerging as a new semiconductor cooling solution.
Due to the limitations of liquid cooling, immersion cooling, which involves submerging semiconductors in immersion cooling liquids to cool them down, has been considered; however, by utilizing MACS to overcome the limitations of liquid cooling, immersion cooling is unnecessary. I believe it is still premature to commercialize immersion cooling. If electronic devices are treated in oil for 5-10 years, unforeseen issues such as softening effects may arise. Additionally, cooling the oil after it has heated up also requires further cooling. I think it will be difficult to implement immersion cooling in practice.
What kind of collaborations are you doing with domestic chip designers?
We supplied heat sinks for Warboy, the core product of FuriosaAI, and secured heat sink supply for the second-generation Renegade by competing with over ten companies. We signed a heat sink development cooperation contract with Rebellion and supplied it to the neural processing unit (NPU) Atom. The heat sinks supplied to Warboy, Renegade, and Atom were air-cooled and did not apply MACS technology. The heat output of the products was not significant, so they were supplied with air cooling; however, as the performance of AI semiconductors improves and heat output increases, liquid cooling will become necessary.
Cooling technology can be extended to other fields.
Almost no electronic products do not use a cooler. Until now, cooling technology has been stagnant; however, a transition is happening comparable to the shift from wired to wireless telephones. In fact, since 2017, MHS has supplied heat sinks for ultrasound diagnostic models from Siemens Healthineers worth about 500 million won each year. We developed a noise-free, environmentally friendly air conditioner without an extruder or refrigerant using MACS technology and completed a patent registration in June last year. By using MHS’s cooling technology, the cooling efficiency of vehicle radiators can be increased by about 30%. In the field of electric vehicles that do not use internal combustion engines, liquid cooling systems are also necessary for motors, inverters, and converters. Almost all machines and electronic devices incorporate cooling systems. Recently, we have been conducting technical consultations with defense companies related to liquid cooling for induction weapons.
MHS’s problem-solving capabilities in the field of cooling technology are unparalleled. We provide technical consulting to various corporations in a multitude of industries, including AI semiconductors, bio and medical devices, aviation, satellites, defense, telecommunications, autonomous driving, electric vehicle circuitry, and batteries.