At this moment in history, what topic is generating more buzz than AI?
Its possibilities—and its possible implications for privacy, equity and labor—have appeared on the research horizon like a blurry new world. While the drive to explore AI and develop responsible applications is high, the computing power required to support the evolving technology is absolutely massive.
Traditional fabrication facilities use a lot of energy and water, and they produce a ton of electronic waste. That level of consumption, said Assistant Professor of Electrical and Computer Engineering (ECE) Tania Roy, is unsustainable.
“Right now, circuitry is so energy-hungry that we usually have to access remote servers, which look like power plants. They have to be cooled like power plants, too,” said Roy, who develops new types of semiconductor devices for AI applications.
Roy cited work by researchers at UC Riverside, which stated that Google’s U.S. data centers used 12.7 billion liters of fresh water in 2021—most of it clean and drinkable. The same amount of water could have produced 5.7 million electric vehicles, according to the authors.
New, more compact hardware would use less energy and occupy a smaller physical footprint. That’s Part I of the research Roy is proposing in her project titled “Independent Neural Network-Enabled Recyclable AI (INNER-I).”
Part II deals with the actual processes used to make the new hardware, which need to be both affordable and recyclable to be sustainable in the long term. Her co-PI and Duke ECE colleague Yiran Chen will co-design this future generation of efficient devices with another Duke ECE co-PI, Aaron Franklin, who is pioneering printable, fully recyclable semiconductor technology. Also collaborating on the project is Duke chemistry professor Michael Therien, who has worked with Franklin to develop the special water-based inks used in his printable transistors.
The desired result? Printed, recyclable devices that use neuromorphic circuitry—whose efficient design is inspired is inspired by the human brain and is quick enough to meet the demands of AI applications— along with an initial framework for an AI system enabled by these devices. “Duke will become the hub where recyclable AI circuitry is made,” predicted Roy. She’s aiming to have a prototype device completed by her new center by the end of its first year.
“There are nearly limitless possibilities for printing neuromorphic devices from nanomaterial-based inks, which opens the way for more environmentally sustainable technologies,” said Franklin. “We will be working through the many options to determine which nanomaterials, printing conditions, and device designs yield the needed performance for enabling neuromorphic technologies.”
The idea holds promise not just for sustainability, but accessibility, too.
“The reason why we have so few faculty focusing on devices in the U.S. is that the cost is exorbitant,” said Roy. “We need access to cleanrooms, sophisticated equipment for characterizing the devices we make. If we have a way to manufacture high-quality tabletop electronics, then we can make the research and development of these devices accessible to everybody.”