A new AI chip could potentially increase power efficiency by sixfold, align computation and data storage in a similar way to biological neural networks and significantly reduce the electrical footprint of AI.

A researcher from Oregon State University’s College of Engineering has helped develop a new artificial intelligence chip that is six times more energy efficient than the current industry standard.

As the use of artificial intelligence increases rapidly, the amount of energy it requires also increases. Projections show that artificial intelligence will account for half a percent of global energy consumption by 2027, consuming as much energy as the entire country of the Netherlands each year.

Seung Chae, an associate professor in the Department of Electrical Engineering and Computer Science, is working to help reduce electricity emissions with this technology. He is investigating microcircuits based on a new material platform that enable both computation and data storage by mimicking how biological neural networks manage information storage and processing.

The results of his research were recently published in a journal. Nature Electronics.

Effective AI processing

“With the emergence of artificial intelligence, computers have to quickly process and store large amounts of data,” Che said. “AI chips are designed to compute tasks in memory; This minimizes data movement between memory and processor; so they can perform AI tasks with higher energy efficiency.”

The chips contain components called memristors, short for memory resistors. Most memristors are made from a simple two-element material system, but the ones in this study feature a new material system known as entropy-stabilized oxides (ESOs). More than half a dozen elements make up the ESO, allowing you to fine-tune their memory capabilities.

Memristors are similar to biological neural networks in that neither has an external memory source, so energy is not wasted moving data from inside to outside and back. By optimizing the ESO composition that is best suited for specific AI tasks, ESO-based chips could perform tasks with much lower power consumption than a computer’s central processing unit, Chae said.

Another implication is that neural networks can process time-dependent information, such as audio and video data, by adjusting the composition of the ESO so that the device can operate on different time scales.