Credit: Pixabay/CC0 Public Domain
The
rapid development and widespread use of artificial intelligence (AI) systems is
posing new challenges for electricity consumption. This is because most AI
systems rely on data centers, facilities hosting several computing servers that
store and process large amounts of data.
Data centers require significant amounts
of electrical power, which makes it harder for providers to offer an affordable
and stable supply of energy across surrounding geographical regions. Most
proposed solutions for meeting this growing energy demand rely on new
infrastructure or additional energy storage systems, which are often expensive
and difficult to deploy rapidly on a large scale.
Researchers and engineers at Emerald AI,
in collaboration with NVIDIA Corporation, Oracle, Salt River Project (SRP), and
the Electric Power Research Institute (EPRI), recently introduced a new
software-based approach to stabilize the power grid by treating data centers as
valuable "flexible" resources. This approach, outlined in a paper in Nature
Energy, entails adjusting the power usage of AI data centers in response to
power grid signals.
"Our paper was motivated by the growing tension between the electric power grid and the rapidly increasing power demands of large-scale AI data centers, which now face multi-year delays in grid interconnection," said Ayse Coskun, Chief Scientist at Emerald AI and co-author of the paper. "Rather than relying on new grid infrastructure, such as additional power generation or transmission, to expand AI capacity, our work explores how data centers themselves can operate as flexible, grid-aware loads."
Power prediction of the platform’s
simulator vs measured power during a power-reduction event. Credit: Nature Energy (2025). DOI: 10.1038/s41560-025-01927-1
Adjusting power consumption in response to grid signals
The primary goal of the recent
study by Coskun and her colleagues was to demonstrate that AI systems running
on GPUs can become flexible grid resources, enabling the efficient use of
available power, without compromising data centers' performance agreements. To
test this idea, the researchers used Emerald Conductor, a software control
framework developed at Emerald AI.
"This framework intelligently
adjusts data center power consumption in response to grid signals while still
meeting application performance and service-level agreements (SLAs),"
explained Coskun. "By analyzing power–performance tradeoffs across
different AI workloads, the system selectively modulates jobs that can tolerate
small adjustments; for example, reducing the power of 'flexible' jobs that can
handle slight slowdowns."
As part of their recent study, the team demonstrated their proposed approach on a real cluster of 256 GPUs running AI algorithms located in a data center in Phoenix. Their results were highly promising, as their strategy significantly reduced power consumption during periods of peak electricity demand.
Image illustrating how a
grid-interactive AI data center operates through a software control layer,
while a results figure showing GPU cluster power modulation. Credit: Colangelo
et al.
Specifically, they reported a reduction in power usage by 25% over a 3-hour period in which the power
grid was under significant stress. All this was achieved with minimal impact on
users, which essentially means that AI systems retained good performance and
could complete tasks both correctly and on time.
Increasing sustainability while supporting AI advancement
Coskun and her colleagues were the
first to demonstrate the potential of utilizing AI data centers to stabilize
power grids in a real-world setting. Their work shows that the smart and
flexible scheduling of AI tasks in response to grid signals can reduce power
usage at times of peak consumption in the grid, while maintaining the quality
of AI-based services and platforms.
"Our work moves the concept of
data center demand response from simulations and small-scale prototypes into
operational reality," said Coskun. "Practically, it sets the stage
for AI data centers to connect to the grid sooner, make more efficient use of
existing capacity, and actively support grid reliability."
In the future, this recent study
could inspire the development of additional strategies that leverage the
capabilities of AI to better manage energy resources. Meanwhile, the
researchers are working on improving their approach and testing it at other data
centers, with the goal of eventually deploying it on a large scale.
"We are expanding this work
through follow-on demonstrations and deployments, deeper integration with GPU
platforms, and ongoing collaboration with industry and grid stakeholders,"
added Coskun.
"Current efforts with partners such as Oracle, NVIDIA, power utilities, ISOs, and EPRI's DCFlex program aim to validate and scale these capabilities under real-world grid conditions. In the longer term, we are exploring coordinated, grid-aware operations across multiple real-world data centers."
No comments:
Post a Comment