Data Centers Need To Go Modular To Cope With AI

The advent of AI presents a significant challenge to the data center industry, which needs to adapt to the demands of a world where servers must be capable of performing vast volumes of computations with massive amounts of data.

With the volumes of computations comes massive amounts of data, which also needs to be stored.

The demand for more servers running on higher-performance chips will also require vast amounts of power, posing fresh challenges in sustainability and green IT.

The investment required will be on a massive scale. Still, according to Chris Sharp, the chief technology officer at Digital Realty, there is no reason to panic.

In a recent blog, Sharp points out that AI-ready infrastructure "isn't just about building one data center.” 

“It requires a modular global data center platform that can support your needs at scale,” Sharp said.

He acknowledges the unique infrastructure requirements of AI as it pushes a new generation of chips, network equipment and other infrastructure into the data center.

“AI requires more computing, network, and storage infrastructure than any shift on the internet since the cloud, and in many cases, we believe it will be even more intense,” Sharp said.  

The Digital Realty response to the change is modularity, which Sharp says is the “key to a timeless design approach.”  

“Digital Realty data centers are comprised of many individual systems, down to the cooling technology used for one rack, on one floor, in one suite, in one building, and on one of our campuses,” he said.

“By designing the data center as modular, interconnected  systems, we can evolve just one or a small number of those systems at a time without needing to redesign or retrofit everything at once.”

This "system of systems" design approach, he says, is cost-effective and operationally efficient and benefits customers because it reduces the time and cost required by them to support new infrastructure.

From a vendor perspective, it also enables a customer's bespoke requirements to be met without impacting others.

“This is vital to meeting the needs of AI, as new deployments are often needed quickly and have specialized needs when it comes to power consumption and cooling,” said Sharp.

For example, some AI infrastructure will draw ten times the industry average power per rack but may still need to be deployed in the same physical space as other, lower energy usage infrastructure. A modular approach easily enables this level of flexibility.

Power hungry

Power consumption due to the need for cooling is a crucial issue for future data centers, particularly as the industry embraces sustainability and Green IT.

Modularity, says Sharp, enables the deployment of sustainable cooling technologies such as liquid cooling.

This cooling method differs from traditional air cooling in that it uses a loop of coolant pumped around the facility, connected to the heat-generating chips in the racks. Then, it moves that heat to a large radiator where it can be removed.

“AI requires more computing, network, and storage infrastructure than any shift on the internet since the cloud.”

Liquid cooling is considered a more efficient way of cooling AI infrastructure because the liquid is 800 times denser than air, making it much easier for the liquid to move excess heat from IT equipment. It also enables data centers to use power from utilities more efficiently, reducing overall power usage.

"Liquid cooling is a notable example of our modular design philosophy," said Sharp.

“Even five years ago, using liquid cooling in the data center was seen as esoteric outside of edge cases – the infrastructure used by the applications of the day simply didn’t generate enough heat to warrant its use. Today, it’s increasingly seen as a hard requirement for dense AI infrastructure deployments.”

“Many of Digital Realty’s data centers have a chilled liquid system that can be used for this purpose, and some don’t—but this doesn’t stop us from providing a liquid cooling solution for our customers.”

High-efficiency air cooling and other alternate liquid cooling technologies can be fitted to a floor, a suite, or even an individual rack.

“We can accommodate these alternatives because our designs are modular,” said Sharp.

Lachlan Colquhoun is the Australia and New Zealand correspondent for CDOTrends and the NextGenConnectivity editor. He remains fascinated with how businesses reinvent themselves through digital technology to solve existing issues and change their business models. You can reach him at [email protected].

Image credit: iStockphoto/KuznetsovDmitry