Researchers from the University of Massachusetts Amherst and their collaborators have demonstrated that their analog computing device, called a memristor, can perform complex scientific computations while overcoming the limitations of digital computing.
This advance could lead to faster and more energy-efficient computing for a wide range of scientific applications, from nanoscale material modeling to large-scale climate science.
Digital computing systems are reaching their limits in terms of speed, energy consumption, and infrastructure when it comes to solving complex equations. Qiangfei Xia, a professor of electrical and computer engineering at UMass Amherst, explains that traditional computing methods require moving data between memory and computing units, which can cause processing “traffic jams” when dealing with large amounts of data.
The team’s solution is in-memory computing using analog memristor technology. A memristor is an electrical component that combines memory and resistance, allowing it to control the flow of electrical current in a circuit while “remembering” its prior state even when the power is turned off. This capability enables the memristor to be programmed into multiple resistance levels, increasing the information density in a single cell.
When organized into a crossbar array, a memristive circuit performs analog computing using physical laws in a massively parallel fashion, significantly accelerating matrix operations, which are frequently used but power-hungry computations in neural networks. By performing the computing at the device site, the need for moving data between memory and processing is eliminated, reducing traffic and increasing efficiency.
Previously, the researchers demonstrated that their memristor could complete low-precision computing tasks, such as machine learning, analog signal processing, radiofrequency sensing, and hardware security. In this new work, they pushed the boundaries further, showing that the technology is also suitable for high-precision scientific computing.
For their proof-of-principle demonstration, the memristor solved static and time-evolving partial differential equations, Navier-Stokes equations, and magnetohydrodynamics problems. Xia emphasizes that it took over a decade for the UMass Amherst team and collaborators to design a proper memristor device and build sizeable circuits and computer chips for analog in-memory computing.
“Our research in the past decade has made analog memristor a viable technology. It is time to move such a great technology into the semiconductor industry to benefit the broad AI hardware community,” says Xia.
#MemristorTech #AnalogComputing #ScientificComputing #InMemoryComputing
