A GPU model does not tell you whether a server needs liquid cooling. The OEM chassis, accelerator power setting, GPU count, fans, rack density, and facility design determine the cooling plan.
NVIDIA says an HGX B200 baseboard can use advanced air or liquid cooling. A liquid-cooled GB200 NVL72 rack is a different product. Ask for the exact server model before planning facility work.
Air cooling
Air-cooled servers use fans and room air. Check the OEM's required inlet temperature, airflow direction, rack spacing, and maximum server power. Eight high-power GPUs can create a large heat load even when the system is certified for air cooling. Your colocation provider may limit power or cooling per rack.
Direct-to-chip liquid cooling
Cold plates transfer heat from selected components to a coolant loop. The design may require a coolant distribution unit, facility water connection, leak detection, service clearance, and separate air cooling for components that are not on the loop.
For a liquid-cooled quote, get the OEM's heat-rejection split between liquid and air. Also request coolant temperature, pressure, flow, connection type, redundancy, and maintenance requirements. A server with cold plates is not a complete facility cooling design.
Immersion cooling
Immersion places compatible equipment in dielectric fluid. It changes the service model and hardware qualification. Confirm OEM warranty and support for the exact configuration before considering it.
Questions to ask the supplier and facility
- What is the full OEM server model and its maximum power draw?
- Is the quoted build air-cooled or liquid-cooled?
- What are the inlet air or coolant requirements at the proposed power setting?
- How much heat remains in the room air?
- What rack power, weight, and service clearance does the system need?
- Who supplies and maintains the cooling equipment?
- Does the quoted warranty cover this installation and cooling method?
Request the full system and facility specifications in writing. Cooling costs vary too much by site to use a generic per-rack retrofit number as a purchase estimate.
Discuss a GPU server build | Browse hardware
