Microsoft/LiquidStack

Companies Develop Immersion-Cooled AI Servers

LiquidStack, a BitFury subsidiary specializing in the development of immersion LSS, and Microsoft presented a standardized two-phase liquid-cooled immersion solution at the OCP Global Summit. The Wiwynn demo servers, designed to OCP Open Accelerator Infrastructure (OAI) specifications, were loaded into a LiquidStack DataTank 4U capable of dissipating approximately 3 kW / 1U (126 kW equivalent per rack).

LiquidStack said it was the first to optimize OCP OAI servers for immersion cooling. According to the company, its DataTank two-phase immersion LSS provides the most efficient solution that requires little space and uses less energy than other cooling systems, while increasing component density and performance.

The demo servers used Intel Habana Gaudi AI accelerators immersed in a liquid dielectric (3M) with a low boiling point, which more efficiently removes heat due to phase transition. This solution should help achieve PUE in the range of 1.02 to 1.03, as it consumes practically no energy. LiquidStack claims its LSS is about 16 times the heat dissipation efficiency of typical air-cooled systems.

According to the company, high-performance computing systems (HPC) already consume a lot of energy to be cooled by air. Therefore, you should use boards designed specifically for liquid cooling, and not try to retrofit those that were created with an air system in mind – when using LSS, the computing infrastructure can take up up to 60% less space.

LiquidStack and Wiwynn are partners. Wiwynn invested $10 million in LiquidStack in April. Wiwynn also tested water-cooled submersible servers at a Microsoft Azure data center in Quincy, Washington. Microsoft, like other hyperscalers, has been exploring the possibilities of immersive LSS for some time now.

AVnews