MHS Develops 'AI GPU Cooling Plate' That Cuts Coolant Use by 60%

"Ranks Among Global Leaders in NVIDIA B300-Class Tests"

MHS, a company specializing in thermal management for artificial intelligence (AI) data centers, announced on October 20 that it has developed an ultra-lightweight cold plate (a GPU cooling plate) that efficiently cools high-performance AI graphics processing units (GPUs) using less coolant.

Image of an AI GPU microchannel cold plate developed by MHS. MHS

As heat output per AI GPU has recently exceeded 1 kilowatt (kW), data centers are rapidly shifting from air cooling to water cooling (liquid cooling), in which coolant flows directly over chips. Because dozens to hundreds of GPUs share a single coolant supply system, optimizing the amount of coolant supplied to each GPU has emerged as a key challenge that determines the efficiency of an entire data center.

MHS said it conducted in-house tests simulating the heat output conditions of an NVIDIA B300-class GPU (1,180 watts (W)), and that its performance metrics surpassed those of leading global products. With coolant flowing at 2.5 liters (L) per minute, the cold plate recorded a thermal resistance of 0.0094 degrees Celsius per watt (°C/W), a measure of the temperature difference per watt. This is lower than the approximately 0.01°C/W recorded by top-tier global products whose results have been publicly disclosed, meaning it rapidly transfers heat away from the GPU.

Even when the coolant flow rate was reduced by 60% to 1.0 L per minute, thermal resistance remained at approximately 0.0124°C/W. This means the cold plate handled a higher thermal load with less coolant than in previously disclosed tests of GB300-class products, which recorded 0.0130–0.0133°C/W at 1,067 W and a flow rate of 1.7 L per minute.

MHS's new cold plate weighs only about 110 grams. Its core heat-exchange section, which transfers heat, is made approximately 1 millimeter thick, reducing the physical load on the GPU and motherboard. It also maximizes space efficiency to suit the needs of densely configured AI servers.

MHS is currently having the product's performance revalidated by an external company specializing in data center cooling equipment. The company plans to further improve cooling efficiency by applying an additional technology called "multilayer microchannels," which dramatically increases the number of microchannels.

"The competition to cool AI data centers is no longer about how cold you can make them, but how little coolant and pumping power you can use to cool them," said MHS CEO Lim Jongsu. "We will prove our technology through comparative tests under identical conditions against leading global products and step up our efforts to enter overseas AI server and data center markets."

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