Cooling Strategy: Air, Liquid-to-Air and Liquid-to-Liquid

Air, liquid-to-air and liquid-to-liquid cooling compared at three densities, with the operational cost and failure modes that follow from each.

At 20kW a rack

Air cooling is adequate and is the simplest answer. It deploys into an existing hall with no coolant loop, no CDU and no additional failure domain. There is rarely a good reason to introduce liquid at this density.

At 40kW a rack

Air becomes difficult and hall-dependent. Liquid-to-air with a rack-level CDU brings cold plates to the servers without requiring facility water, which makes it deployable in space that was never plumbed. The consequence must be stated plainly: the heat is rejected into the room, so room cooling has to absorb it. Cooling that was adequate for ordinary servers will not be, and the load moves from the rack to the CRAC units rather than disappearing.

At 80kW a rack and above

Liquid-to-liquid is the only practical option. A CDU separates the facility loop from the technology loop, holding the server side at the temperature, pressure and water quality the hardware requires while the facility side can be whatever the building provides. Coupling them directly is how corrosion and fouling problems begin.

Operational cost

In this climate cooling is a substantial fraction of facility power, so the choice affects operating cost and not only capital cost. Warm water cooling deserves specific attention: at inlet temperatures high enough for ambient rejection it can avoid a chiller entirely, removing both a large continuous electrical load and a system that can fail.

Failure modes

Air fails gradually and visibly through rising temperatures. Liquid fails faster and needs leak detection, but N+1 pumping makes a CDU maintainable with the row still running, which air-side maintenance at high density often is not. Neither is inherently more reliable; they fail differently and should be operated accordingly.