The approach encodes input data as temperature distributions and exploits heat flow through engineered materials to carry out matrix-vector multiplications, fundamental operations for machine learning applications. Experiments demonstrated over 99% accuracy in these calculations using excess heat within electronic devices rather than traditional electric current.
Although still in early development stages with challenges remaining for scaling to complex deep-learning models, this method could enhance energy efficiency and reduce chip space by eliminating multiple temperature sensors. As reported, researchers view heat as an informational resource rather than waste, but broader applicability and integration into commercial computing remain to be established.
Source & further reading
MIT engineers design structures that compute with heat ↗Source published 2026-01-29. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.