Chips, compute and data centres

Liquid cooling

In short

Liquid cooling removes heat from servers by circulating a liquid, rather than using air alone. A cold plate or an immersion bath takes heat off the equipment, and a separate loop can pass that heat to the facility cooling system.

Also called
Liquid-cooled computing, Direct liquid cooling
Topic
Chips, compute and data centres
Level
Beginner

Liquid cooling carries heat away in a liquid. In cold-plate cooling the liquid stays in a closed loop, and cold plates on hot components conduct heat into it. The Open Compute Project separates that from immersion, where a dielectric liquid touches the components, and from indirect cooling, where air carries heat to a liquid. A coolant distribution unit isolates the equipment loop from the facility loop and moves heat through a heat exchanger1. The US Department of Energy describes that path, from a closed water loop into the condenser-water loop rather than the room air, and says water can carry significantly more heat than air, with pumps typically more efficient than fans2.

Cold-plate liquids include water with additives, glycol mixtures, dielectric fluids and refrigerants. Single-phase liquids stay liquid. Two-phase liquids boil and then condense. The fluid has to suit the materials it touches1. Immersion instead submerges equipment in a dielectric liquid, which conducts heat and does not conduct electricity. Single-phase immersion keeps that liquid liquid. Two-phase immersion boils it and condenses the vapour. Fluids named for immersion include synthetic hydrocarbons, esters and fluorochemicals3.

There is no general chip-power level that requires liquid cooling. The Open Compute Project says it is more efficient than air, but the choice depends on density, cost and whether air can still meet temperature limits1. One hybrid puts cold plates on the hottest parts and fans on the rest. Nvidia's DGX GB racks do that: manifolds and cold plates cool the Grace CPUs and Blackwell GPUs, and fans cool networking and storage4.

Example

In an Nvidia DGX GB rack, liquid flows through manifolds and cold plates on the CPUs and GPUs. Networking and storage in the same trays are still cooled by fans, and the guide includes leak detection to limit damage to the equipment and the building systems around it4.

Why it matters

The US Department of Energy says water can carry significantly more heat than air, and that pumps are typically more efficient than fans. Cold-plate designs add a separate liquid loop, a heat exchanger and leak management12.

Questions people ask

Is liquid cooling the same as immersion cooling?

No. Immersion puts a dielectric liquid in direct contact with the components. Cold-plate cooling keeps the liquid inside plates and pipes attached to those components. The Open Compute Project treats the two as different methods13.

Does liquid cooling use water?

Often, but not only water. Cold-plate guidance lists water with additives, glycol-based liquids, dielectric liquids and refrigerants. Immersion uses a dielectric liquid, because that liquid touches the electronics13.

Can liquid-cooled servers leak?

The designs assume they might. Cold-plate guidance calls for leak detection and a response plan. Nvidia's DGX GB rack guide includes leak detection to prevent damage to the equipment and secondary damage to the infrastructure14.

Related terms

Sources

  1. Open Compute Project, Liquid Cooling Cold Plate Requirements, revision 1.0
  2. US Department of Energy, Cooling Water Efficiency Opportunities for Federal Data Centers
  3. Open Compute Project, Immersion Requirements, revision 2.10
  4. NVIDIA, DGX GB Rack Scale Systems User Guide: Hardware

Back to the glossary