Coolant Distribution Unit Selection for Liquid-Cooled AI …
Quick Summary
- Function: Pumps coolant, manages temperature, rejects heat to facility loop
- Capacity: CDUs range from 30kW to 2MW+ heat rejection
- Coolant Types: Water-glycol, dielectric fluids, refrigerant-based
- Redundancy: N+1 CDU configuration recommended for AI clusters
- Integration: CDU connects to facility chilled water or dry cooler system
Coolant Distribution Unit Liquid-cooled AI servers for AI Liquid Cooling
Coolant Distribution Units (CDUs) are the heart of any liquid-cooled GPU infrastructure, responsible for circulating coolant, managing temperature, and transferring heat from server-level cooling loops to facility-level heat rejection systems. Selecting the right CDU is critical for AI cluster reliability, efficiency, and scalability.
CDU Types and Technologies
| CDU Type | Capacity Range | Coolant | Best For |
|---|---|---|---|
| Rack-Mounted CDU | 30-80 kW | Water-glycol | Single rack deployments |
| Row-Level CDU | 100-500 kW | Water-glycol / dielectric | 2-10 rack clusters |
| Facility-Level CDU | 500 kW - 2 MW+ | Water-glycol | Large AI data centers |
| Rear Door HX | 20-50 kW per rack | Facility chilled water | Hybrid air-liquid cooling |
| Immersion CDU | 50-500 kW | Dielectric fluid | Immersion-cooled deployments |
Key Selection Criteria
Cooling capacity must match or exceed the total thermal load of connected GPU servers with 20-30% safety margin. For a rack of 8x H100 servers consuming 56kW, an 80kW CDU provides adequate capacity with margin. Pump redundancy is essential—N+1 pump configuration within the CDU ensures continuous operation during pump maintenance or failure.
Temperature Management
CDUs control coolant temperature through heat exchanger bypass valves and variable-speed pumps. The industry standard approach maintains server inlet coolant at 25-35°C for direct-to-chip cooling, enabling year-round water-side economization in most climates without chiller operation. Lower coolant temperatures (15-20°C) provide additional thermal margin but reduce economizer hours and increase pump energy consumption.
Integration with Facility Systems
CDUs connect to facility-level heat rejection through a secondary coolant loop connected to dry coolers, cooling towers, or chillers. For AI clusters exceeding 500kW, the facility loop design significantly affects CDU effectiveness. NTS provides comprehensive CDU integration services including facility connection design, coolant chemistry management, and monitoring system integration.
Related Content
Explore more about this topic:
- Liquid Cooling vs Air Cooling for AI Racks
- Data Center Tier Classification for AI
- Multi-Modal AI Model Infrastructure
How many servers can one CDU support?
A typical 80kW rack CDU supports 8-12 H100 GPU servers depending on GPU configuration and workload. A 500kW row CDU supports 50-75 servers. Facility-level CDUs can support 200+ servers in a modular deployment.
What coolant is used in GPU liquid cooling?
Standard coolant is deionized water with 20-30% propylene glycol for freeze protection and biocide for biological growth prevention. Dielectric fluids are used in immersion cooling where electrical conductivity is a concern.