Analog computing can solve complex equations and use far less energy

Analog computing can solve complex equations and use far less energy
by Clarence Oxford
Los Angeles CA (SPX) Mar 19, 2024
Researchers at the University of Massachusetts Amherst, alongside global partners, have demonstrated that their analog computing device, known as a memristor, is capable of executing complex scientific computations more efficiently than conventional digital computing systems. This discovery has significant implications for fields ranging from climate science to material science, where complex equations are foundational.

Professor Qiangfei Xia of UMass Amherst, a key author of the study published in Science, highlighted the limitations of current digital computing, which suffers from inefficiencies related to the transfer of data between memory and computing units, likening the process to a "traffic jam." In contrast, the team's memristor-based approach integrates memory and computing into a single unit, dramatically reducing data transfer needs and enhancing efficiency.

Memristors, devices that combine the functions of memory storage and resistance, offer a promising alternative by retaining information even when powered off, unlike conventional transistor-based chips. These devices enable higher information density and can be programmed to exhibit multiple resistance levels, thereby facilitating in-memory computing. The researchers' novel circuit architecture and programming protocol allow for the representation of high-precision numbers through a combination of multiple, low-precision analog devices like memristors, reducing circuit complexity, energy consumption, and latency. This advancement extends the applicability of memristor technology from previously demonstrated low-precision tasks, such as machine learning, to high-precision scientific computing.

The study successfully demonstrated the memristor's capability to solve intricate equations, including static and time-evolving partial differential equations, by leveraging the memristor's ability to perform computations directly at the device level, akin to working remotely to avoid traffic congestion.

This achievement marks the culmination of over a decade of research by the UMass Amherst team and their collaborators, paving the way for the integration of analog memristor technology into the semiconductor industry, with potential widespread benefits for AI hardware and beyond.

Research Report:Programming memristor arrays with arbitrarily high precision for analog computing