New computer chips show promise for reducing energy footprint of artificial intelligence

CORVALLIS, Ore. – An Oregon State University College of Engineering researcher has helped develop a new artificial intelligence chip that could improve energy efficiency six times over the current industry standard.

As the use of artificial intelligence soars, so does the amount of energy it requires. Projections show artificial intelligence accounting for half a percent of global energy consumption by 2027 – using as much energy annually as the entire country of the Netherlands.

Sieun Chae, assistant professor of electrical engineering and computer science, is working to help shrink the technology’s electricity footprint. She is researching chips, based on a novel material platform, that allow for both computation and data storage, mimicking the way biological neural networks handle information storage and processing.

Findings from her research were recently published in Nature Electronics.

“With the emergence of AI, computers are forced to rapidly process and store large amounts of data,” Chae said. “AI chips are designed to compute tasks in memory, which minimizes the shuttling of data between memory and processor; thus, they can perform AI tasks more energy efficiently.”

The chips feature components called memristors – short for memory resistors. Most memristors are made from a simple material system composed of two elements, but the ones in this study feature a new material system known as entropy-stabilized oxides, or ESOs. More than a half-dozen elements comprise the ESOs, allowing their memory capabilities to be finely tuned.

Memristors are similar to biological neural networks in that neither has an external memory source – thus no energy is lost to moving data from the inside to the outside and back. By optimizing the ESO composition that works best for specific AI jobs, ESO-based chips can perform tasks with far less energy than a computer’s central processing unit, Chae said.

Another upshot is that artificial neural networks would be able to process information that’s time dependent, such as data for audio and video, thanks to tuning the ESOs’ composition so the device can work on a varied time scale.

Funded by the National Science Foundation, the study was led by researchers at the University of Michigan; Chae participated as a doctoral student at Michigan before joining the faculty at Oregon State.

The collaboration also included researchers from the University of Oklahoma, Cornell University and Pennsylvania State University.