Data centre energy impact, defined by AI

Data centres have been in the firing line to reduce their impact on grid energy use for some time now. Prior to 2020, the industry focused on network availability, with energy efficiency often being a secondary consideration in facility planning and operations. But since the pandemic — when our appetite for cloud adoption, streaming and video conference calls saw worldwide network traffic surge by more than 40% — the industry’s carbon impact greatly increased.

The latest generative AI systems rolling out are far more compute intensive than older versions of the technology. While the industry spent the better part of the last few years building resilience and scalability into the infrastructure that make our digital lives possible, there is work ahead to support energy efficient data centres in the era of high compute, defined by AI.

Readying an AI model can create the same level of carbon emissions as five cars throughout their lifetimes, according to MIT Technology Review. Tirias Research forecasts that by 2028 an additional 4250 megawatts (MW) of data centre power will be required to drive emerging AI systems, a 212x increase over 2023. While daunting to consider, this is both a challenge and opportunity for the years ahead.

On one hand, as datasets become more complex and models become ever more clever, emissions will increase. But AI holds a golden opportunity to apply the very same technology causing the reset to data centres themselves.

For example, AI giant Nvidia, the near-trillion-dollar company at the centre of AI’s power, is teaming up with Intel to focus on reducing the energy consumption of the server, with the goal to run workloads 25 times more efficiently than predecessor systems. Alongside Vertiv and other industry experts, Nvidia also secured US$5 million to develop an innovative direct liquid cooling and immersion cooling system, which is projected to operate at 20% higher efficiency than conventional approaches.

It’s logical for Gartner to estimate we will see half of all cloud data centres leveraging AI by 2025. Without industry innovations of this type, data centre operators will sweat over working to balance running a data centre at optimum efficiency, while maintaining resilience. The analytics are often large, the implications for failure are catastrophic, and it is all for a small, yet continuous saving.

AI can analyse vast amounts of real-time energy usage data and monitoring environmental factors, able to make fine-tuned adjustments to primary power and cooling systems.

As AI becomes more common in the data centre, we expect to see it capture savings that stack up over the years. And leading the charge will be the large hyperscale and colocation operators who have set goals to become carbon neutral or carbon negative by the end of this decade.

Hyperscale and colocation data centres are miles ahead of inhouse computing rooms when it comes to efficiency strategies. Their scale allows them to advance technologies to combat emissions, such as liquid immersion and direct-to-chip methods, far more quickly.

But as AI’s prevalence in enterprise grows, we’ll also likely see latency requirements force users to rearchitect and apply edge data centres nearby to build the ecosystem beyond colos.

The entire ecosystem has an important role in developing a roadmap for the industry — a roadmap that will illustrate how AI can allow data centres to operate like well-oiled machines, delivering high performance without an exorbitant environmental cost.

It’s going to be a productive few years for industry developing one of the fastest-growing technologies, rolling it out, and doing so in a way that optimises power use and lowers operational expense.

Image credit: iStock.com/BlackJack3D