As GPUs and AI accelerators push beyond one kilowatt of power consumption, many systems builders are turning to liquid cooling to manage the heat. However, these systems still rely on complex networks of plumbing, manifolds, and coolant distribution units (CDUs) to make it all work.

On Thursday, LiquidStack, best known for its immersion cooling tech, launched a monster CDU with more than one megawatt of cooling capacity that it claims can plug into any existing platform using direct-to-chip liquid cooling.

CDUs are responsible for pumping the coolant to and from connected systems and racks exchanging the captured heat via a liquid-to-air, or in this case, liquid-to-liquid heat exchanger.

While a megawatt of capacity might seem like an obscene amount of cooling, when it comes to cooling modern AI systems, it doesn't go as far as you might think. Next-gen systems from companies like Nvidia will cram somewhere between 5.4kW and 5.7kW of compute in a single RU server and upwards of 120kW in a single rack.

At 1.35 megawatts of cooling capacity, LiquidStack's first CDU has enough capacity to cool an entire Nvidia DGX GB200 Superpod with its 288 2,700W Grace-Blackwell Superchips with room to spare. If you're curious about that system, you can find our deep dive here.

Rent-a-GPU outfits like CoreWeave and Lambda are deploying tens of thousands of accelerators. Even with a megawatt cooling capacity, they will still need a lot of CDUs if they want to take advantage of Nvidia's most compute dense configurations.

One caveat to high-capacity CDUs, like LiquidStack's, is their potential blast radius in the event of a failure. To mitigate some of this risk, they claim their CDU offers N+1 redundancy via hot-swappable pumps, which can be equipped with redundant temperature, pressure, and flow monitor and expansion tanks if desired.

LiquidStack's entrance into the CDU space comes as datacenter operators scramble to capitalize on the AI hype while it lasts, with shortages of critical components like rear-door heat exchangers and CDUs being a particular pain point. LiquidStack clearly aims to capitalize on this pent up demand. However, they're far from the only one.

The extreme demand has driven some purveyors of liquid cooling infrastructure, like Vertiv, to massively expand production. Earlier this year, the firm revealed plans to ramp up the production of liquid cooling equipment 45-fold by the end of 2024. ®

https://www.theregister.com//2024/08/22/liquidstack_cdu_ai/