Pioneering Research on Memristors Pushes the Frontier of Artificial Intelligence Hardware
Content
“Noise,” or troublesome fluctuations in electrical conductance, is a major factor limiting the precision of a memristor, a potentially game-changing electrical component that can store information without power in many computational applications for machine learning and artificial intelligence. Now Professor Qiangfei Xia of the UMass Amherst Electrical and Computer Engineering Department and leading experts from TetraMem Inc., the University of Southern California, and the Massachusetts Institute of Technology have solved the noise problem in memristors with some avant-garde research described in a paper published in the distinguished scientific journal Nature.
As the researchers who contributed to the Nature paper explain quite succinctly, “We devised a process to eliminate the noise in individual memristors.” The research team has thus blazed a pathway for fully integrating memristive devices into the semiconductor technology that underlies much of today’s computing circuitry.
By “denoising” memristors, the research project demonstrates a record-high number of conductance levels, or levels at which energy can flow easily through memristors, to a remarkable 2,048 as achieved in memristors in large arrays and fabricated in a chip factory.
The researchers note that the 2,048 conductance levels are “as far as we know, a record number among all types of reported memory devices.”
The memristor is a two-terminal electrical component that controls the flow of electrical current in a circuit, while also “remembering” the prior state, even with the power turned off. The non-volatile memristor is so important because it can boost throughput and energy efficiency for machine learning and artificial intelligence (AI) through parallel analog computing by using physical laws, especially in “edge applications” (referring to the modern, distributed computing architecture that brings data storage and computation closer to the data source).
Xia is a pioneer in integrating emerging devices with silicon-based CMOS (complementary metal-oxide-semiconductor) chips, taking advantage of both the novel properties of emerging devices and the maturity of silicon technology for machine intelligence applications. As demonstrated by his team’s approximately 30 high-impact papers published in Nature family journals in recent years, the hybrid chips, previously made by integrating foundry-made silicon transistors with memristors at the UMass Lab, are best suited for low-power and high throughput edge AI applications.
“The results reported in the Nature paper were obtained from a wafer with both the CMOS circuitry and memristors made together in a commercial foundry, marking the successful lab-to-fab transfer of memristor technology,” as Xia explains. “Furthermore, this work means memristors can also be used for applications beyond edge AI, such as scientific computing, that requires a much higher precision.”
Nature’s Chief Applied & Physical Sciences Editor Karl Ziemelis expresses the significance of the groundbreaking research in a “Research Briefing” piece this way: “The number of conductance levels that have been reliably addressed in this system is striking in its own right. But when you add to that the insights gained into the origin – and, ultimately, the control – of the underlying memristive switching process, and that this was achieved for devices prepared in a standard semiconductor foundry, the academic appeal and potential practical value of this paper become obvious.”
As the researchers explain, “Our approach and conclusions are generally applicable to all memristive materials, including oxides, nitrides, sulfides, and carbides, although the specific material phases involved might differ for different oxide systems. Encouragingly, we were able to make these high-precision memory devices both in a university laboratory and in a commercial chip factory.”
As the researchers conclude, “We have demonstrated that robust, uniform, high-precision memristors in large arrays can be fully integrated on the complementary metal-oxide-semiconductor technology that underlies much of today’s computing circuitry. Our finding thereby creates the potential for large-scale commercial applications. The hope is that this potential can soon be realized by mainstream chip manufacturers.” (March 2023)