Memristors for AI Can AI chips get a sense of time?

Related Vendors

University of Michigan

FAULHABER Drive Systems

Wuxi InfiMotion Technology Co., Ltd.

Time is measured in the brain through neurons that relax at varying rates after receiving a signal. Similarly, memristors—hardware analogs of neurons—can replicate this process.

Artificial neural networks could soon be able to process time-dependent information such as audio and video data more efficiently than before. In a study led by the University of Michigan (U-M), the first memristor with an adjustable "relaxation time" is reported in "Nature Electronics".

Memristors, electrical components that store information in their electrical resistance, could reduce the energy requirements of artificial intelligence by around 90 times compared to today's graphical processing units. It is already predicted that AI will account for around half a percent of total global electricity consumption in 2027, and this figure could increase as more companies sell and deploy AI tools.

Gallery

"There is a lot of interest in AI right now, but to process bigger and more interesting data, the network needs to be scaled up. That's not very efficient," said Professor Wei Lu of U-M along with John Heron, U-M Associate Professor of Materials Science and Engineering and co-author of the study.

Energy guzzler GPU

The problem is that graphics processors work very differently to the artificial neural networks on which the AI algorithms run. The entire network and all its interactions have to be loaded one after the other from external memory, which consumes both time and energy. In contrast, memristors offer energy savings because they mimic important aspects of how artificial and biological neural networks work without external memory. In a sense, the memristor network can embody the artificial neural network.

"We anticipate that our brand-new material system could improve the energy efficiency of AI chips by six times compared to the state-of-the-art material without varying time constants," said Sieun Chae, a recent U-M PhD materials scientist and co-author of the study with Sangmin Yoo, who also received his PhD from U-M.

Neurons and memristors

In a biological, neuronal network, time measurement is achieved through relaxation. Each neuron receives electrical signals and sends them on, but there is no guarantee that a signal will be passed on. A certain threshold of incoming signals must be reached before the neuron sends its own, and within a certain amount of time. If too much time passes, the neuron relaxes as the electrical energy escapes from it. Neurons with different relaxation times in our neural networks help us to understand event sequences.

Memristors work a little differently. It is not about the presence or absence of a signal, but about how much of the electrical signal is allowed through. When a signal is present, the resistance of the memristor decreases, allowing more of the next signal to pass through. In the case of memristors, relaxation means that the resistance increases again over time.

While Lu's group had investigated how relaxation time could be incorporated into memristors in the past, this was not something that could be systematically controlled. Now, however, Lu and Heron's team have shown that variations of a base material can provide different relaxation times, allowing memristor networks to mimic this timing mechanism.

Superconductor in play

The team based the materials on the superconductor YBCO, which consists of yttrium, barium, carbon and oxygen. Although it has no electrical resistance at temperatures below -292 °C, they wanted to use it because of its crystal structure. It guided the organization of the magnesium, cobalt, nickel, copper and zinc oxides in the memristor material.

Heron calls this type of oxide an entropy-stabilized oxide, the "kitchen sink of the atomic world": the more elements they add, the more stable it becomes. By changing the ratio of these oxides, the team achieved time constants of 159 to 278 ns. The simple memristor network they developed learned to recognize the sounds of the numbers zero to nine. Once trained, it was able to recognize each number before the audio input was completed.

These memristors were produced using an energy-intensive process because the team needed perfect crystals in order to measure their properties accurately. Simpler manufacturing processes are expected for series production. "So far it's a vision, but I think there are ways to make these materials scalable and affordable," Heron said, "because they're abundant in the earth, non-toxic, cheap and you can almost spray them on."

This research was funded by the National Science Foundation. It was conducted in collaboration with researchers from the University of Oklahoma, Cornell University and Pennsylvania State University. The device was fabricated at the Lurie Nanofabrication Facility and studied at the Michigan Center for Materials Characterization. (sb)