A New Era in Computing
For decades, computing has followed a fundamental rule: data is stored in one place and processed in another. This separation, the foundation of von Neumann architecture, has driven the digital age. Today, a new player is challenging this established order: the memristor, a device capable of both storing and processing data simultaneously. This is the line of research being pursued by Francisco Rivadulla and Ramón y Cajal Diogo Costa, researchers at the Singular Centre for Research in Biological Chemistry and Molecular Materials (CiQUS). Costa will lead his own memristor project in collaboration with CiTIUS, as part of the USC-Televés Chair.
How Do These Components Work?
Memristors can store information because their electrical resistance changes based on the current previously passed through them. "It’s a resistance with memory," Rivadulla explains. In traditional computing, for example, information is stored in binary format, using 0s and 1s depending on whether the resistance is low or high. Memristors, however, function similarly but with a key advantage: "They can adopt not just two states, but multiple resistance levels," explains Diogo Costa.
For instance, if a memristor can be configured into four resistance levels, it can represent 00, 01, 10, or 11 in a single device. This way, with the same number of physical components, it is possible to significantly increase memory density. Additionally, memristors can perform calculations such as matrix multiplications, which are essential for artificial intelligence.
Instead of constantly sending data between memory and processor, memristors allow computation to occur directly where the data is stored. "This enables faster data processing," explains the project leader. This concept is known as in-memory computing, as it combines storage and processing in a single device.
Transistor vs. Memristor: What’s the Difference?
The most basic component in electronics is the transistor, whose primary function is to store information in the form of 1s and 0s. It operates based on two states: 0 (off) and 1 (on). Transistors rely on semiconductors, and their state is not preserved when power is cut off; additional circuits are required to maintain memory.
In contrast, a memristor can represent multiple states (00, 01, 10, 11, etc.), while a transistor can only represent one (0 or 1). This increases the density of information that can fit on a single chip. "They are non-volatile memory elements and enable new computing architectures," explains Francisco Rivadulla.
However, the true revolution lies in the concept of in-memory computing. In a classical computer, data constantly travels between memory and processor, consuming time and energy. Memristors combine these two functions in one place, resulting in multiple advantages: less data transfer, lower energy consumption, and greater speed.
Goals: Optimization and Application
The project launched by these researchers from the CiQUS Materials Science Group aims to improve the materials that make up memristors and demonstrate their usefulness in real-world artificial intelligence tasks. The work begins with the basics: "The operation is based on the movement of ions within the material," says Francisco Rivadulla. In many devices, the ions that move are oxygen, and the materials used are oxides.
The first objective is to optimize the resistive response of these memristive materials: modifying the chemical composition, adjusting the structure, and improving the stability of their resistance states. The second objective is to "implement a functional demonstrator" using a matrix of memristors, Rivadulla notes.
This demonstrator will be used to perform AI tasks directly on hardware. It will be used to identify patterns, such as recognizing handwritten numbers, classifying images, and, in the future, distinguishing objects in an image, such as separating a person from an animal. "We’re starting with numbers because it’s the simplest task," the researcher clarifies.
Creating a Prototype
The prototype being developed at CiQUS is a proof of concept. Diogo Teixeira explains that the initial idea is to use these numbers, made up of lit and unlit segments, and check if the memristor matrix can identify, for example, a 7, 8, or 3. The next step will be to increase the difficulty: enabling the hardware to recognize handwritten numbers. "So it can distinguish between a 9 written by me and a 9 written by you," explains Ramón y Cajal researcher Costa.
Starting from the memristive material, usually an oxide, a device is built using lithography—creating microscopic patterns on a chip, depositing the necessary layers, and fabricating the structure that will act as a memristor. This process transforms the material into a functional device. Once the chip exists, an image like a digital number can be converted into an electrical signal, and the resistive response allows for recognition. "We handle microfabrication and electrical control of small chips," Costa explains.
However, they are not yet at the industrial application stage: thousands of memristors are needed to achieve this, and CiQUS lacks the necessary materials. "We want to create a demonstrator that shows the functionality and advantages of our material," clarifies the project leader.
The Future of Computing
In today’s digital landscape, memristors are attracting research attention because they overcome one of the main limitations of traditional transistor-based computing: "You can have more memory in the same space," explains researcher Diogo Costa.
With memristors, the separation between storage and data processing disappears. Additionally, they can perform operations like matrix multiplications, which are the foundation of many artificial intelligence algorithms, including language models like ChatGPT.
"It avoids the need to have two separate locations for each part," he summarizes, describing in-memory computing. From CiQUS, they are paving the way to revolutionize computing from Galicia.
By Andrea Veiga, GCiencia.
Investigadores do CiQUS e CiTIUS establecen unha liña de traballo sobre os memristores