从风扇到液体：人工智能数据中心对液冷技术的迫切需求.pdf

1、From Fans to FluidOCP Global Summit 2025Bobby KinstleSr.Thermal&Power Engineer,Hyve SolutionsThe Critical Need For Liquid Cooling in AI Datacenters High energy costs and reduced system density Fan power consumed up to 50%of total system energyAir Cooling ChallengesAir cooling hit a wall 30KWPER RACK

2、AI workloads changed everything2 MWPER RACK BY 2030Rack power expected to continue to increase as high asAI Power Trending Ever Upwards1000W+to 500W+TDPCPU power roseGPU power demandsjumped from 300W toRack power density increased from 30KW to 100KW+AI Power Requires Liquid Cooling Previously consid

3、ered too expensive and risky Now essential for high-density AI infrastructureLiquid cooling is finallymainstream CPU racks now can enjoy the benefits of liquid cooling Facility cooling systems get a much-needed efficiency boost through the use of warm waterLiquid Cooling is the Density Answer Smalle

4、r cold plates allow unprecedented hardware density Higher CPU and GPU density shortens cable runs 1200W Chip Cooling is easy 2500W Superchip MCM modules Thermal uniformity improvements unlocks amazing potentials for 10,000+W wafer scale AI accelerators CPU Racks fully populated again.Lets go 56U!Liq

5、uid Cooling is the Design Answer Temperature sensitive devices can be located anywhere in the system without worry of upstream preheating.Heated liquid leaves the high-density computer for easy transport outside to large efficient radiators and coolers.Dont think side car,think outside the building.

6、Hot liquid has opportunities for useful energy recovery for things like facility heating and melting ice from sidewalks and roadways in the winter.Liquid Cooling is the Performance AnswerBase Frequency with TDP HeadroomAll Core Turbo ActivatedMax Turbo Activat