From the CTO’s corner: Simon Gardner, Oper8 Global Group

We wrote recently that NVIDIA high-performance computing at 600kW per rack is coming; a five-fold increase in the maximum densities we are seeing today. One of the implications of this is that advanced liquid cooling infrastructure will be essential to deal with the heat being generated by these racks.

Single-phase and two-phase direct-to-chip (D2C) liquid cooling are two evolving methods to manage the thermal demands in high-density, high-power facilities – not just for AI factories but also for HPC data centres of any size.

There are multiple advantages of D2C liquid cooling in general; most notably, using liquids with high thermal conductivity (such as water, dielectric fluids, or engineered coolants) is up to 1,000 times more efficient than air cooling. That in turn means reduced energy consumption, a lower environmental impact, cutting back on the space needed for the data centre, and improved overall server room performance.

In the case of two-phase D2C cooling, there is the added benefit that the two-phase process, where the liquid is allowed to boil and convert to gas, absorbs more heat at lower flow rates as compared to a single-phase solution. That’s because the conversion from liquid to gas requires a lot of energy, which is drawn from the heat generated in the racks, and the gas itself acts as a coolant, reducing the need for other infrastructure (plus energy and water!) to cool down the liquid before it circulates back through the chips.

In the drive towards net zero data centres – where the lower the energy and water consumption the better – D2C liquid cooling plays an important role. However, there are a number of challenges in implementing these systems.

Both single-phase and two-phase systems are complex to install, especially if you have to retrofit racks with cold-plate plumbing and manifold connections – and two-phase is additionally complex because you have to manage the coolant in two states; liquid and gas.

Retrofitting existing facilities with D2C cooling may even be impossible, depending on the data centre space, existing fire suppression systems, and cabling and piping layout. If that’s the case, a  micro or modular approach is the one we’d recommend, because you can build all the additional infrastructure into the prefabricated unit before installing into the space, which is what we did with a leading international tyre manufacturer.

At the moment, there are also high capital and operational costs to implement D2C cooling. Cold plates, pumps and specialised tubing need to be installed, plus, in the case of single-phase, coolant distribution units (CDUs). Managing and monitoring those fluids becomes more critical, because leaks can be far more damaging, even catastrophic to data centre operations.

While both systems reduce the environmental footprint of your data centre, some dielectric fluids used for two-phase cooling pose health and environmental risks, attracting additional regulatory scrutiny and compliance and worsening the impact of any leaks.

On the plus side, D2C liquid cooling allows us to deliver far greater densities in the data centre, and reduce white space by up to 60% – a huge cost saving when you consider that the  typical land value in some locations is doubling, due to competition for space from some of the big data centre operators.

If you are building a greenfield site, that might help to assuage the upfront costs from implementing D2C systems. Alternatively, if you are looking for leased facilities, the reduced space won’t just reduce your rent; it might also give you far greater options for real estate in more convenient locations with better access to services and resources.

So, where to from here? It’s critical that you consider the cooling implications from increasing your data centre density, and that you also keep an eye on the future. Ironically, we are starting to see AI helping to sort out the problems of its own creation.

A good example of this future is Microsoft’s new ‘microfluidic cooling system’, which the vendor is claiming to provide up to three times better cooling performance than cold plates. The technology is using AI to help optimise a bio-inspired design to cool chips’ hot spots. It takes the heat signatures of chips to identify hotspots, then etches liquid coolant channels onto the back of the chips.

Of course, this is just in the prototype stage, but these technical innovations will continue at pace as the industry comes up with new ways to keep our data centres cool. In preparing for D2C cooling, and keeping your eye on the future, it’s critical to stay nimble and agile enough to take advantage of any groundbreaking solutions as they come to market.