Understanding AI's resource needs and sustainability challenges
Discover how AI's demands are affecting data centers' power and water resources, cooling challenges and the approaches being implemented to address these concerns.
As AI transforms the data center industry, its growing resource demands pose significant challenges. The surge in AI capabilities drives higher power consumption and cooling needs, underscoring the urgency of adopting sustainable energy.
This article examines the balance between harnessing AI's potential and managing its impact on vital resources while highlighting current challenges and innovative strategies to address them.
Concerns of increased energy consumption
GPUs are the heart of AI technology. GPUs, with their thousands of small cores, are designed to handle massive parallel processing workloads, making them ideal for large-scale AI training. However, GPUs consume 200 watts to 400 watts each.
CPUs have fewer cores, run sequential processing and draw power in the range of 65 watts to 250 watts. GPUs are technically more efficient than CPUs when measured in floating-point operations per second (FLOPS). GPUs are reported to achieve 50 FLOPS/Watt in some studies, compared with 10 FLOPS/Watt for CPUs.
The enormous processing demands of AI require large numbers of GPUs, resulting in very high total power demand during AI "training" mode. When training reverts to low-power "communications" mode, power drops abruptly. Power spikes back up when training resumes, and the cycle continues.
The trend is for AI to push these numbers even higher, with predictions of 1 MW racks and Terawatt-scale data centers, which are beyond the generation and transmission capabilities of almost all localities.
The impact on our power infrastructure is huge. A few years ago, power levels of 80 kW to 150 kW per rack were considered eye-opening, and total data center loads of 10 MW to 50 MW were considered high-end. Today, we see 250 kW to 500 kW, half a Megawatt, per rack and total data center loads in the Gigawatts, or thousands of Megawatts.
The trend is for AI to push these numbers even higher, with predictions of 1 MW racks and Terawatt-scale data centers, which are beyond the generation and transmission capabilities of almost all localities. Power "swings" can reach 10 MW in milliseconds.
Concerns of water scarcity
GPUs are liquid-cooled, but their heat must be transported outside the facility. Water has historically been the medium for this because it's several thousand times more efficient than air at removing heat. But that efficiency comes at the expense of evaporation, and huge data centers can vaporize thousands of gallons of water per hour. That's made water usage the most challenging issue the industry is tackling today. There are ways to generate more power, but water is a finite resource, and, in many climates, it is simply scarce. Unfortunately, some of those climates are otherwise great locations for large data centers.
The Green Grid developed a metric called Water Usage Efficiency (WUE). It's similar to the well-known PUE metric, but, like PUE, it applies only to operating facilities. It does not enable authorities to set or enforce usage standards, particularly in advance of construction, so addressing water usage has become a priority for every Standards and Code committee.
Recognizing that water demand can lead to project denial, most large new data centers have reverted to air cooling. This solves the water-use problem but at the expense of significantly higher power demand. Newer magnetic-bearing air-cooled chillers deliver greater efficiency, but the overall power demand without water cooling remains substantially higher.
Concerns over increased noise pollution
A significant amount of effort is focused on addressing the power and cooling requirements of AI, but another important concern is noise pollution. The numerous air-cooled units used in these facilities can generate substantial noise, hindering acceptance by local municipalities.
Most jurisdictions have noise regulations in place, but complying with these limits can be challenging. For instance, sound barriers designed to reduce noise may obstruct airflow, thereby reducing cooling efficiency. The support and infrastructure for AI development still have a long way to go, especially in a rapidly changing environment.
Offset grid strain with power plants
The most efficient approach is for facilities to bring their own power. Large new facilities are either constructing their own power plants or buying and reactivating old ones. The Three Mile Island nuclear power plant, set to open in 2027 to meet the demands of Microsoft data centers and the grid, is an example.
Still, most decommissioned plants are, unfortunately, coal- or gas-fired. That increases greenhouse gas emissions. In the best of circumstances, new data centers will build wind or solar plants, but these have variable capacity and only supplement local generation. Building fuel cells at waste dumps that emit sufficient methane would be ideal, but it requires very careful site selection.
Managing DCIM software and power spikes
Power spikes are another challenge. Whereas conventional data center power demand changes only a few percent over seconds or minutes, the instantaneous swings in AI facilities are enormous. Power infrastructures are not designed for this. Consider the instability at home when power is restored to a neighborhood after a regional outage. Now imagine this happening and destabilizing the entire power grid many times a day. AI data centers require special power designs to negate these effects.
AI is also being used, in essence, to manage itself. DCIM software packages, now a necessary tool for operators to keep control of large, complex facilities, all incorporate some form of AI.
Managing and optimizing power and cooling in real time is one feature, with AI used to anticipate transient loads. That helps, but a more complete solution would add bidirectional UPS. Using newer active front-end rectifiers and fuzzy logic or model predictive control, these systems not only maintain reliable power to the servers but also maintain a constant battery charge, allowing them to return power to the grid when voltage drops. This stabilizes the grid and can also reduce energy costs by peak shaving and assisting the grid during peak demand. Due to the large number of discharge/recharge cycles, lithium-ion batteries are proving to be the battery of choice in these systems. However, newer technologies, such as ultracapacitors, are being explored.
Robert McFarlane is senior principal in charge of data center design for the international consulting firm Shen Milsom and Wilke LLC. McFarlane has spent more than 40 years in communications consulting and has experience in every segment of the data center industry.
