MIT researchers have developed an analog computing method that uses waste heat from electronic devices as a form of data input rather than discarding it. Microscopic silicon structures, designed via a physics-based optimization algorithm, route heat flows to perform matrix-vector multiplication — the core operation in machine learning models — with over 99% accuracy in simple cases. While scaling to modern deep-learning workloads remains challenging, the technique has near-term potential for on-chip thermal sensing without additional energy consumption or dedicated temperature sensors.
Sort: