I was just running the audio video in the discussion hall, but cooling data centers was a common topic when I worked for Giant Tech Firm ~20 years ago. The engineers always said the easiest way to do it was just make the room bigger, but actually doing that was never going to fly with upper management. Seeing “unused” space was simply untenable.
Another idea I heard was building the things fairly high up on the windward slopes of mountain ranges like the Cascades or Sierra Nevada. Partially underground, with unpowered ventilation systems that use the constant upwelling wind.
But, no. We have to pack it all tight and build them in places that aren’t naturally cold.
If you’re building it somewhere with cheap land (and why wouldn’t you?), it seems like “just make the room bigger” would usually be the cheapest option.
Modern data centres, especially AI systems have really high density. It isn’t just a desire for less real estate. Although that helps. AI systems are operating as a parallel computer and latency matters. In addition to simple speed of light issues, every time you step to the next level of interconnect bandwidth slows down. You really want high density.
Fluid cooled systems have always been around. CDC, Cray, in the ‘60s and ‘70s. But air was always cheaper, and upstart supercomputer companies like Thinking Machines made a point of air cooling being cheaper. Back then 10s of kilowatts was a lot, and say 5 kilowatts per rack the sort of ballpark.
Now power density is vastly higher, and getting the heat out exceeds anything air can reasonably manage.
But that water is in a closed loop, and isn’t a consumable.
When you are trying to move heat about you are going to find it impossible to ignore the fabulous utility present in the latent heat of evaporation of water. Water takes a lot of energy to evaporate and that is energy you suck out of your data centre. Depending upon how warm you are happy to run your systems you might manage with only that.
If you need to run things colder, heat pumps come in, but even here, the colder they see the ejection end, the higher their coefficient of performance. And that directly affects the energy needed to run the data centre. You can see a doubling of CoP adding a water cooling tower. That might mean a 10-15% drop in your data centre’s electricity use. If you are already power limited, this is impossible to ignore.
« We have also launched a new AI datacenter design that uses a closed-loop system. By constantly recirculating a cooling liquid, we can dramatically cut our water usage. In this next-generation design, already deployed in locations such as Wisconsin and Georgia, potable water is no longer needed for cooling, reducing pressure on local freshwater systems. »
Bolding mine. They acknowledge using potable water but only the amount it takes to initially fill the closed system.