Memristor-based hardware to advance AI

Artificial intelligence (AI) is rapidly emerging as a force in nearly every sector of society. Yet the computer hardware in use today is based on hardware introduced over 75 years ago, relying on the transistor device first invented in 1947.

Thus far, computer engineers have largely kept up with the increased demands of advancing technology through various ‘brute force’ methods, such as decreasing device size and increasing bandwidth, explains Qiangfei Xia, the Dev and Linda Gupta Professor of Electrical and Computer Engineering at UMass Amherst. The result, though, is that AI programs are enormously expensive to run, both in terms of monetary cost — limiting access to major corporations and the super wealthy — and environmental impacts, including carbon emissions and freshwater use. For example, training a large language model that powers the popular chatbot ChatGPT could cost over $10 million and consume more than 700,000 litres of fresh water.

“AI is incredible. In 2016, we saw its potential when AlphaGo, an AI-based computer system created by a Google subsidiary, beat the reigning human champion at the ancient Chinese board game Go,” Xia said. “But what most people don’t know is that the computers running AlphaGo filled nearly an entire room, and playing one game cost thousands of dollars in electricity.”

For the past decade, Xia and his collaborators have been developing a ‘memristor’ device to build new computers. They have demonstrated that this new analog computing device can complete complex computing tasks while bypassing the limitations of digital computing and using far less energy. And they believe that memristor technology holds the potential to advance AI and address many of today’s most pressing scientific questions, from nanoscale material modelling to large-scale climate science.

The Nanodevices and Integrated System Lab, headed by Xia, has been addressing pressing issues in AI hardware and has a significant impact on emerging hardware based on transition metal oxide memristive devices. Xia has been recognised for his contributions to this field, as an elected Fellow of the Institute of Electrical and Electronics Engineers (IEEE).

The promise of memristive computing

According to Xia, the concept of memristor devices dates back to 1971, when it was first proposed by a professor at the University of California Berkeley. Work on the memristor remained mostly theoretical for decades, but in the late 2000s, HP — where Xia was working at the time — made a breakthrough by connecting the memristor concept to a physical device the company built in the lab.

While transistors rely on the movement of electrons, the memristor takes inspiration from the human brain, in which ions move through hundreds of trillions of synapses to carry information. The memristor controls the flow of electrical current in a circuit, while also ‘remembering’ the prior state, even when the power is turned off, unlike today’s transistor-based computer chips, which can only hold information while there is power. In the memristor, computing is performed at the site where data is stored, rather than moving data between the computer’s memory and processing modules.

Xia draws an analogy between this form of ‘in-memory computing’ and the empty roads during the early days of the COVID-19 pandemic. “Everyone was working from home, so that reduced traffic on the roads substantially.” For a computer built with memristive technology, these empty ‘roads’ mean a huge boost in energy efficiency and computing throughput. This opens many doors for creating low-power AI hardware, especially for edge computing, where data is processed in the devices that collect it, rather than being sent to a centralised cloud server. Potential applications include consumer electronics — such as lightweight AR/VR goggles or wireless earbuds — scientific research and military technology.

Xia and his collaborators see great potential for commercialising this technology. In 2018, they founded TetraMem, a Silicon Valley-based startup of which Xia is a co-founder and advisor.

“We want to work together and transfer this technology to mainstream foundries so it can benefit users more broadly in the future. The memristor technology is now far enough along in development that this is an ideal time for industry to take over. In university research, we’ll continue fine-tuning the technology and exploring novel applications, such as in 6G cellular network technology and language processing, to name a few,” Xia said.

Computing inspired by the human brain

Going forward, Xia aspires to build computer circuits that are even more inspired by neuroscience and work as efficiently as the human brain. “We want to design a memristor device that’s a step closer to how our biological neurons work. This is known as neuromorphic computing,” Xia said.

Returning to the story of AlphaGo’s victory, Xia said the computer was estimated to use around 150 kilowatts of power to play the game, while its human competitor used only about 20–25 watts in its highly efficient brain. The human brain is also incredibly powerful in its capabilities. For example, we can recognise another person instantaneously based on only partial information, while digital computers must still do pixel-by-pixel matching.

“We have a long way to go. Literally, we do not understand our brains well enough yet. We want to use what we have already learned from the brain to build the next-generation computer. This will require collaboration with not just electrical engineers and computer scientists but also neuroscientists and psychologists. I feel so lucky to be working in this field during this time when AI is booming. I’m very hopeful for the next phase of truly brain-inspired computer hardware,” Xia said.

This is a modified version of a news item published by the University of Massachusetts Amherst. The original version of the news item can be accessed here.

Image credit: iStock.com/WANAN YOSSINGKUM