MHS, a company specializing in thermal management for artificial intelligence data centers, said on the 20th that it has developed an ultralight GPU cold plate that cools Nvidia graphics processing units with less coolant. The technology is seen as easing cooling bottlenecks at AI data centers because it allows more GPUs to run on the same cooling equipment.
In tests based on the heat load of Nvidia's Blackwell B300-class GPUs, or 1,180 watts, MHS recorded thermal resistance of 0.0094 degrees Celsius per watt when coolant flowed at 2.5 liters per minute. That is lower than other top-tier global products, which stand at around 0.01 degrees Celsius per watt. A lower thermal resistance figure means heat is removed more quickly. In effect, the plate handled a higher heat load with less coolant.
The cold plate weighs only about 110 grams. The company made the core heat-exchange section, which transfers heat, about 1 millimeter thick to reduce the load on the GPU and the motherboard.
Based on these results, MHS plans to seek external performance verification through a company specializing in data center cooling equipment. "The competition in AI data center cooling is now less about how cold you can make things and more about how little coolant and pumping power you use to do it," MHS Chief Executive Lim Jong-soo said. "We will prove our technology through head-to-head comparison tests with top global products under identical conditions and move aggressively into overseas AI server and data center markets."