Three data center hubs in France and the United States will be equipped with experimental methods to make them more flexible and adapt to the needs of their local power grids. IEEE spectrum Learned.
The selected hub will serve as a test bed for solutions to rising demand for AI power and is expected to double in the next five years, accounting for 3% of total global consumption. These forecasts raise concerns about fully meeting the power demand for data centers without sacrificing the reliability of everyone else, and they drive global efforts to innovate around the issue.
The experimental hub originated from a groundbreaking collaboration called the DCFLEX Initiative, consisting of grid operators, utilities and large tech companies including Google, Meta, Micta, Microsoft, Nvidia, and Oracle. The groupElectric Power Research Institute (EPRI) in Palo Alto, California, announced its locations for its first three locations today and is expected to announce up to seven sites this year.
“It’s like the first cohort,” said Anuja Ratnayake, EPRI’s emerging technology executive. “These three are marriages between the available marriages our members are involved in and some use cases that require the longest testing period.”
AI workload orchestration
The selected hubs are operable, grid-connected data centers, each of which will test different aspects of flexibility. In Lenor, North Carolina, Google will work directly with local utility Duke Energy to arrange and move the computing workload of data centers to meet the needs of the grid.
This approach is called workload orchestration, and involves a data center offloading its compute tasks to other facilities or pushing these tasks to different time frames to reduce the overall load at a given time. Ratnayake said the strategy only targets large ratings like Google, which use data centers for AI training and operate on schedule.
Ratnayake said that unlike AI training, enterprises and cloud services that provide streaming and online banking often cannot transfer loads in time and geographic fashion, as these services satisfy requests in real time to people on the other side of the screen. “If it’s a banking service or a credit card service, then there’s little flexibility because it’s something that handles transactions at a very high quantity in a very short window,” she said.
That’s where the other two sites come in. At the second site, two data center workloads in Phoenix provide a mix of AI training and cloud services from Nvidia and Oracle. Third-party solution provider Emerald AI will coordinate orchestrate with local utilities including the Salt River project.
Earlier this year, the Phoenix site simulated a peak event where energy demand on the grid is high to test whether the data center can respond by reducing its workload. Ratnayake said the test was successful and the website achieved 10% to 40% flexibility in orchestration workload. Next, the team will test whether the site can respond to a true peak energy event.
Power stability of UPS system
The third site in Paris is committed to maintaining the stability of the data center during power outages. Most large data centers are equipped with uninterrupted power supply (UPS), a backup system that keeps the situation running when power is created from the grid. Data center operator Data4 will work with French transmission system operator Schneider Electric and RTE to explore how to use the UPS system to power through voltage and frequency issues. Currently, voltage drops and other grid problems can trigger data centers to offline to protect their computers from damage.
There may also be opportunities to use energy storage from UPS systems to provide additional power flexibility to the data center. The DCFLEX team will explore this at future sites, along with data center cooling strategies and low-carbon alternatives to diesel-powered backup generators.
The team plans to conduct experiments at up to 10 sites this year. These locations will be distributed in the United States and Europe, with one or two likely in the Middle East or Asia. So far, DCFLEX has accumulated 45 collaborators, which can be carried out from 14 launched in October 2024. By co-operating these field tests, the team aims to build a flexible framework for flexibility that supports AI-driven load growth.
EPRI expects results from the first three demo sites later this year.
From your website article
Related articles around the Internet
