🤖 AI & Beyond

Four Reasons to Feel Positive About AI’s Energy Consumption

3/ More Efficient Cooling in Data Centers
Another significant source of energy demand is the need to manage the waste heat generated by the advanced hardware on which AI models operate. Tom Earp, engineering director at the design firm Page, has been constructing data centers since 2006, including a six-year period working for a major tech company. Earp seeks efficiencies in everything from the structure of the building to the electrical supply, the cooling systems, and the methods of data transfer in and out.

For over a decade, as Moore’s Law has slowed, data-center designs remained largely consistent, says Earp. However, a shift has occurred. With the transition to processors like GPUs and newer chip designs on the horizon, predicting the hardware requirements for a new data center—and the associated energy demands—has become challenging. In the short term, one can safely assume that chips will continue to increase in both speed and heat output. “What I see is that those making these decisions are preparing for significant increases in power requirements,” states Earp.

One fact stands out: The chips powering AI models, such as GPUs, demand more power per unit of space than older types of computer chips, leading to substantial implications for a data center’s cooling infrastructure. “When power goes up, heat goes up,” notes Earp.

With so many high-performance chips packed together, traditional air cooling (big fans, essentially) is inadequate. Water has emerged as the preferred coolant since it excels at dissipating heat. This can pose challenges for local water sources near data centers; however, there are strategies to enhance water-cooling efficiency.

One approach involves using water to transfer waste heat from a data center to locations where it can be repurposed. For instance, Denmark has utilized water from data centers to provide heating for homes. In Paris, during the Olympics, it was employed to heat swimming pools.

Water can also act as a type of energy storage. Renewable energy generated by wind turbines or solar panels can be harnessed to cool water, which can then be stored for later use, helping to reduce power consumption during peak times.

However, as data centers heat up, relying solely on water cooling is insufficient, according to Tony Atti, CEO of Phononic, a startup specializing in cooling chips. Chipmakers are producing chips that transfer data at increasingly rapid rates. Atti highlights Nvidia’s upcoming chip, capable of processing 1.6 terabytes per second: “At that data rate, demand for cooling rises exponentially,” he explains.

According to Atti, the chips within servers consume roughly 45% of a data center’s power. However, cooling those chips now requires almost as much power—around 40%. “For the first time, thermal management is becoming the barrier to the growth of this AI infrastructure,” he concludes.