AI and the chips that run it are so hot right now.

No, like, really hot, physically hot. So hot that Equinix has devised plans to expand support for liquid cooling across 100 datacenters in 45 metros to quench customers' thirst for accelerated computing.

While liquid cooling has been around for decades, it has largely been deployed in high-performance computing environments supported by bespoke kit from the likes of HPE Cray, Evident/Atos, and Lenovo. With the rise of AI, enterprises are looking for places to park large clusters of thermally challenging accelerators.

Chips like Nvidia's H100 or AMD's new MI300X are capable of consuming 700-750W of power a piece under load. It's not just one accelerator either - up to eight of these chips are packed into a single box, and several racks worth of systems may be required to support an enterprise's model requirements.

Equinix has previously supported liquid-to-air cooling through technologies like in-rack heat exchangers. This usually involves a small closed-loop system, including a coolant distribution unit, reservoir, and heat exchanger. Heat generated by liquid-cooled servers is captured and then rejected into the hot aisle.

These closed-loop systems have become popular in older datacenter environments lacking the water necessary to support wide-scale deployment of liquid-cooled systems.

This gave them "the ability to have dense deployments without having direct facility chilled water to their cabinet or cage environment," Tiffany Osias, VP of global colocation at Equinix, explained in an interview with The Register. The downside is their cooling capacity is generally limited and they take up rack space that could otherwise be used by servers.

However, as demand for AI infrastructure has grown - customers are now asking for racks with upwards of 60-80kW capacity - Equinix is looking to make deploying direct liquid-cooled systems more affordable.

For those that aren't familiar, direct liquid cooling passes coolant through cold plates attached to hotspots like CPUs and GPUs - but sometimes also memory and NICs too. Where this heat goes from there varies depending on the infrastructure. For a datacenter with facility water, a coolant distribution unit equipped with a liquid-to-liquid heat exchanger is often employed to reject the heat to cooling units on the exterior of the building.

Beyond being more efficient at capturing and rejecting heat, liquid cooling has other benefits - particularly when it comes to power consumption. As chips have gotten faster and hotter, a significant amount of power is being consumed by fans. According to Equinix, as much as 30 percent of a cabinet's power consumption is directly attributable to fans.

Reducing power consumption by eliminating fans can allow for denser deployments, or lower operating costs. In certain environments, Osias posits that there is also the potential to reuse the heat generated by these facilities for things like district heating.

For those customers bent on deploying GPU systems for AI applications, but unready to commit to direct liquid cooling, Equinix is also expanding the use of rear door heat exchangers. These are rack-sized radiators through which cool facility water flows. As hot air exits a conventional server, it passes through the radiator, where it is cooled back to acceptable levels.

"What's important for datacenter operators is that they maintain flexibility by making their facilities interchangeable between air-cooled cabinets and liquid-cooled cabinets. This ensures that when they need to deploy liquid cooling, they're ready to do so quickly, without requiring significant retrofitting to existing facilities," Equinix explained in a recent blog post.

We've seen other colocation providers like Digital Reality and Colovore adopt this technology in order to support even the densest AI deployments.

In an interview with The Next Platform this summer, Digital Realty CTO Chris Sharp claimed that by using rear door heat exchangers it could support rack densities of up to 90kW. Meanwhile, Colovore says its new datacenter in Santa Clara can manage 250kW racks using direct liquid cooling. ®

https://www.theregister.com//2023/12/13/equinix_liquid_cooling/