• Telco AI Leader: Data center industry leaders like Vertiv and Equinix are supporting the need for more liquid cooled data center infrastructure to meet demands of rising GPU use.
• Dream Bigger: While today hyperscalers are deploying the technology, private AI will drive demand for liquid cooling in smaller enterprise data centers, with potential for bottlenecks.
Data centers around the world hosting consumer and enterprise applications may see a shift towards using liquid cooling techniques, as opposed to the various forms of air cooling (e.g., room, rowm and rack) that are used for the vast majority of data center workloads today. Liquid cooling instead uses water or a refrigerant to dissipate heat from CPUs and GPUs. The advantage of liquid cooling is the higher thermal efficiency of transferring heat through water as opposed to the air. While liquid cooling is not a new technology, it has been eschewed in favor of air cooling due to the more CapEx-intensive nature of building liquid cooling systems within data halls, and the ability of air conditioning systems to largely keep up with the heat generated by IT workloads, up until this point in time.
Signals from industry are indicating that there may soon be a shift towards various liquid cooling systems due to the increased use of GPUs to power high-density compute for workloads like AI model inferencing. This is driving up thermal loads of data centers beyond what many air-cooling systems can manage. For example, Vertiv sees liquid cooling as a core concept for its data center of the future. New hyperscale facilities, purpose built for the likes of AWS, Azure, and Google Cloud Platform are architected with liquid cooling systems in mind, to support the new generation of GPU intensive high-performance compute. However, the company also sees a need for brownfield enterprise level data centers to implement liquid cooling (for more details, please see AI’s Hidden Impact and Vertiv’s Vision for the Data Center of the Future, March 28, 2024). Others in the industry also echo this sentiment. Equinix, by some metrics the largest data center operator in the world, is also seeing demand for liquid cooling, not only from the largest hyperscalers, but also from regular retail enterprise tenants. Equinix, which provides wholesale single or low tenant data centers to hyperscalers, termed xScale, is seeing wide deployment of liquid cooling at these centers already. However, more recently the company is exploring deployment of liquid cooling at its retail IBX data centers. There are liquid cooled halls in Paris and Osaka with more coming in markets like Australia.
Equinix believes the demand for smaller scale data centers with liquid cooling will come from demand for private AI compute services. While the largest deployments of AI workloads currently reside within hyperscale environments, as the market matures many enterprises may require the use of GPUs in their private enterprise data center space. Enterprise datasets used to train and inference models can be massive. This can incur very high costs related to data egress as well as performance as data flows from a private database into a hyperscale environment. Further some enterprise may be concerned with privacy or data sovereignty when dealing with AI workloads. Lastly, latency can be an issue when dealing with large volumes of data, or mission critical applications supported by GPUs. All these factors are driving demand for GPUs not just for hyperscale, but enterprise scale as well. This may create a bottleneck as the industry moves along the maturity curve in the adoption of AI at scale. Not only are GPU shortages still impacting the industry, driving up prices, the data center infrastructure, like liquid cooling, needed to support usage, is not yet in place for smaller scale deployments. The future may be liquid cooled, but the journey has just started.
IT Connection 