Liquid Cooling Loop Design

Overview

Liquid cooling loop design for AI racks involves three heat exchange stages: GPU-to-coolant at the cold plate, primary-to-secondary loop within the CDU, and secondary loop-to-chiller in the data center. Each stage has distinct efficiency trade-offs determined by the inlet water temperature, flow rate, and the temperature delta between GPU junction temperature and coolant.

Inlet water temperature is a key system-level decision. Colder inlet water (e.g., 20°C) creates a larger temperature delta at the GPU cold plate, enabling more heat extraction at lower flow rates and allowing a single CDU to serve more racks — but increases chiller energy. Warmer inlet water (e.g., 40–45°C) reduces chiller load but requires higher flow rates for the same heat removal. A 20°C loop and a 45°C loop are completely different system designs, requiring different CDU specifications, flow controls, and chiller configurations.

Hose material selection affects both routing within the chassis and long-term serviceability. Stainless steel hoses have a smaller outer diameter than EPDM rubber, freeing cable routing channels in constrained 1U chassis, but require stress-distribution shrouds at cold-plate joints to prevent metal fatigue cracking under repeated servicing.

ODMs typically perform L10 (chassis-level) and L11 (rack-level) thermal and hydraulic simulations, collecting boundary conditions such as ambient temperature, cold-aisle temperature, hot-aisle temperature, customer supply water temperature, and data center geography, to determine the appropriate CDU specification and rack-per-CDU count for a given deployment.