Dust-sized silicon chips use waste heat to do math with 99% accuracy

Researchers at MIT have developed a novel silicon chip that performs mathematical computations using waste heat as the medium for data processing, achieving over 99% accuracy. Unlike traditional digital computing that relies on voltage-based logic, this analog computing technique encodes input data as precise temperatures, which flow through specially designed porous silicon structures. The heat distribution resulting from this process represents the computation output, maintained at a fixed temperature. The team demonstrated matrix-vector multiplication—a fundamental operation in machine learning—with high accuracy, showcasing the potential for energy-efficient thermal sensing and signal processing without additional energy costs. The innovation hinges on a new software system employing inverse design, where desired computational functions are defined first, and algorithms iteratively create the optimal silicon structure geometry to conduct heat accordingly. Each structure is about the size of a dust particle and processes data via heat conduction, encoding matrix coefficients in its shape. A challenge arose because heat conduction only allows encoding positive coefficients; the researchers overcame this by splitting matrices into positive and negative parts, processing them separately,