Navigating Liquid Cooling Architectures for Data Centers with AI Workloads
Schneider Electric · Paul Lin, Robert Bunger et al.
A Schneider Electric white paper that describes the ecosystem around liquid-cooled AI servers (heat capture in the server, the coolant distribution unit and heat rejection) and compares six common architectures.
On this page
Overview
The paper splits the ecosystem around liquid-cooled AI servers into three elements: heat capture in the server, the coolant distribution unit (CDU), and how heat is rejected outdoors. CDU functions are temperature control, flow control, pressure control, fluid treatment and filtration, heat exchange and isolation. Heat-exchange types are liquid-to-air (L-A), liquid-to-liquid (L-L) and liquid-to-refrigerant (LR). Typical capacities are given: rack-mounted CDUs about 20 to 40 kW for L-A and about 40 to 80 kW for L-L; floor-mounted CDUs up to about 60 kW for L-A and about 300 kW to more than 1 MW for L-L.
In a liquid-cooled server, 60 to 90% of heat goes to the liquid depending on how many components are cold-plated; the remainder goes to air (CRAH, RDHx). Six architectures are built from CDU type and heat-rejection path: L-A rack and L-A floor CDUs rejecting to room air; L-L rack and L-L floor CDUs rejecting to existing facility water; and L-L rack and L-L floor CDUs rejecting to a dedicated water system.
Architectures that reject heat into room air limit density to about 40 kW per rack, L-L about 80 kW. Pros, cons and when-to-use guidance is given for each. A shared plant creates a trade-off: liquid loops can run about 40 C supply while air cooling needs about 20 C, so a common chiller wastes economizer potential. The paper also stresses that facility chiller water must not flow directly through cold plates.
Key points
- Rack-mounted CDUs: about 20 to 40 kW (L-A) or about 40 to 80 kW (L-L).
- Floor-mounted CDUs: L-A up to about 60 kW; L-L about 300 kW to more than 1 MW.
- 60 to 90% of a liquid-cooled server's heat goes to liquid depending on how many components are cold-plated; the remainder to air (CRAH, RDHx).
- Architectures rejecting heat into room air (L-A CDU) limit density to about 40 kW per rack; L-L about 80 kW.
- Shared plant trade-off: liquid loops can run about 40 C supply while air cooling needs about 20 C, so a common chiller wastes economizer potential.
- CDU key functions: temperature control, flow control, pressure control, fluid treatment and filtration, heat exchange and isolation.
- Six architectures: {L-A rack, L-A floor} rejecting to room air; {L-L rack, L-L floor} to existing facility water; {L-L rack, L-L floor} to a dedicated water system.
Using it in design
- Model CDUs as catalog items with exchange type (LA, LL, LR), form factor (rack, floor), capacity kW, approach temperature, flow L/min and redundancy.
- Validate against the paper's ranges: LA-rack 20 to 40 kW, LL-rack 40 to 80 kW, LA-floor up to 60 kW, LL-floor 300 to 1,000+ kW.
- Build the architecture picker on the six combinations, and block liquid-to-air CDUs for racks above 40 kW.
- Recommend a dedicated high-temperature loop when the liquid-cooled load is about 500 kW or more (see WP 210), and always insert a CDU between facility water and the TCS.
Monthly newsletter · free
Keep your reading list current
Once a month: new library sources, changes to Turkish regulations and upcoming data centre events.
SoSo CAD
Check cooling capacity alongside your design
Review zone heat loads, available capacity and the capacity balance after the largest unit is lost in SoSo CAD.
Request a demo
Who should read it
Mechanical designers and infrastructure planners choosing a liquid-cooling architecture and CDU type for AI clusters.
Related standards
- ASHRAE TC9.9 Liquid Cooling Guidelines (W-classes)
- OCP Advanced Cooling Facilities
- ASHRAE TC9.9 Thermal Guidelines
Publication details
- Authors
- Paul Lin; Robert Bunger; Victor Avelar
- Publisher
- Schneider Electric
- Year
- 2025
- Document number
- WP 133
- Document language
- English
- Source type
- Manufacturer
- Access
- Free to read
- Licence
- Copyright retained; citation only — the publisher’s terms apply to the original document
Citation
Paul Lin; Robert Bunger; Victor Avelar (2025). Navigating Liquid Cooling Architectures for Data Centers with AI Workloads. Schneider Electric. WP 133. https://download.schneider-electric.com/files?p_Doc_Ref=SPD_WP133_EN
Events on this subject
- SC26 – International Conference for High Performance Computing, Networking, Storage and Analysis
- NVIDIA GTC
- ISC High Performance
All content and rights belong to the respective authors and publishers. SoSo CAD provides only the citation, its own summary and a link to the source, and is not responsible for the sources' content.