Liquid Cooling for AI, HPC and High Density Racks

Liquid cooling removes heat from IT equipment using a liquid instead of air. Liquid carries roughly 3 500 times the heat of the same volume of air, so the heat can be taken at or near the components that make it rather than blown across a room. In a data centre it takes three forms: a heat exchanger in the rear door of the rack, cold plates sitting directly on the processors, or the whole server submerged in a dielectric fluid.

Air cooling holds a rack to roughly 30 to 40 kW before the volume of air required becomes impractical to move and contain. Below that, well designed air cooling with proper containment is usually the cheaper answer. Above it, liquid is the only practical route, which is why every current AI training platform assumes it. Rear door cooling covers the 25 to 65 kW band, direct-to-chip and immersion go well beyond it.

They capture heat at different points. A rear door heat exchanger cools the air as it leaves the rack, so nothing inside the server changes and it retrofits into racks you already own. Direct-to-chip puts cold plates on the processors and takes 70 to 80% of the heat straight into the liquid, which is the route rack-scale GPU platforms are designed around. Immersion submerges the whole server in dielectric fluid, removes the air path entirely and gives the highest density per square metre.

Yes. Rear door heat exchangers are the usual retrofit route because they fit standard 600mm racks, add only 310mm of depth on the chilled water versions and need no change to the servers. They can be fitted rack by rack across a series of maintenance windows. Where there is no facility water loop at all, a self-contained DX rear door needs only an outdoor condenser, and a liquid-to-air CDU lets direct-to-chip run without building a chilled water system first.

The cost is driven by the method, the capacity and how much of the loop already exists, not by a per-rack list price. Fitting rear doors to a handful of hot racks on an existing chilled water loop sits at one end of the range; building a new immersion hall with its own heat rejection sits at the other. Send us the rack count, the kW per rack, the platform you are deploying and whether you have a water loop today, and we will configure and price it rather than quote a range.

It is the first question everyone asks, and the answer is not that a loop will never leak. It is that the loop is built so a leak is detected in seconds, isolated automatically and reported before a temperature changes. Every CRX4 coolant distribution unit carries leak detection, level and flow sensing and continuous temperature logging. The secondary loop that touches your servers is physically separated from the facility loop by a heat exchanger, so a problem on one side stays on that side.

A coolant distribution unit sits between your building services and the servers. It keeps the secondary loop that touches the IT physically separate from the primary facility loop, holds supply temperature and flow steady, filters the fluid and reports over Modbus or SNMP. Direct-to-chip and immersion both need one. Choose it by scale: a rack CDU is 4U and serves one cabinet at 100 kW, a row CDU is floor standing and serves a whole aisle at 1 520 kW or 2 020 kW depending on the water it is fed.

Not necessarily, and this is where most of the running cost saving comes from. Direct-to-chip and immersion both run on water far warmer than a traditional chilled water loop, so a dry cooler or a cooling tower can carry the load for most of the Australian year with the compressors off. Free cooling chillers from 100 kW to 1053 kW are available where mechanical cooling is still needed, and they drop from 316 kW of draw to 41.6 kW on the largest unit once free cooling engages.

It can, and it should be checked before anything is ordered. Some server vendors support immersion directly, some support it on specific models, and some do not. Fluid compatibility with gaskets, thermal interface materials, labels and optics also needs confirming, and fans are removed as part of the preparation. We check this against the exact hardware you intend to deploy as part of the design rather than after delivery.

The saving comes from two places: fans that no longer have to move large volumes of air, and heat rejection that can run without mechanical refrigeration for much of the year on warm water. The immersion range quotes a cooling PUE factor of 0.03, which is the energy overhead of the immersion loop itself. That is not a whole-facility PUE, which also carries UPS losses, lighting and any remaining air handling, so treat it as the cooling contribution rather than the number for the building.

Yes. UPS Solutions designs, supplies, installs, commissions and maintains liquid cooling across Australia, New Zealand and the Pacific Islands, with local technicians in Sydney, Melbourne, Brisbane, Perth, Adelaide and beyond. Maintenance covers the loop, fluid condition, filtration and heat rejection on the same 24/7 response as your UPS.