Despite exponential growth in demand for AI, it will not necessarily have as serious an impact on power grids as many think it will.
This is according to Cor van Nugteren, General Manager: Thermal at ISF Group, a diversified engineering, procurement, construction and management (EPCM) company.
Analysts have pointed to AI as a catalyst changing the face of power grids around the world, with some noting that AI will result in data centres tripling their power consumption and putting grids under strain.
Van Nugteren says AI will change data centre design and their power requirements, but the future is not as gloomy as it has been painted.
“In colocation data centres, not all clients will be using AI. I estimate that in the next few years, around 70% of the resources will be allocated to GPUs for AI processing, but the remaining 30% of the resources will remain historical, running standard systems such as corporate mail servers and financial processing,” he says.
“This will drive change in data centre design, with white space becoming smaller. It will also mean more power consumption, with data centres harnessing newer cooling methods such as glycol closed-loop cooling systems and heat exchangers. Due to cost and potential capacity limitations in future, some data centre operators may also move to generate their own power offsite and wheel it over various distribution networks to their data centre facilities.”
Van Nugteren notes the impacts of AI are not dissimilar to those experienced at the advent of supercomputers: “Historically, the Centre for High Performance Computing has done this before. In 2012, its 1 petaFLOP supercomputer required 18 cabinets, and power consumption was around 500kW. Just over 10 years later, to do 1 petaFLOP, you need only one cabinet with power consumption of just 80kW. So the space requirement has reduced significantly and the power consumption has dropped in the order of six times. Despite the exponential growth in AI adoption, there are advances in computing and cooling technology and efficiencies have increased dramatically.”
While AI will certainly bring about changes in data centres, it’s not all bad news, he says. “Units are getting more efficient and using less space, and data centres can now get away with not using active cooling, so reducing their power consumption dramatically.
“The systems we require are becoming much less energy hungry, plus you don't need active water chillers anymore. You can cool systems with more efficient dry coolers and higher water temperatures. Where a power usage effectiveness (PUE) efficiency rating of 1.3 would be considered very good in a data centre today, innovations in cooling to use higher water temperatures result in a PUE of under 1.15. This means a 50% reduction in power consumption for data centre cooling,” he says.
Share