MIT engineers design structures that compute with heat (news.mit.edu)

MIT engineers have innovatively designed silicon structures that leverage excess heat for computation, providing a potentially more energy-efficient alternative to traditional electronic devices that rely on electricity. By encoding input data as temperature variations and employing heat flow through meticulously structured materials, these devices can perform complex calculations, such as matrix vector multiplication, with over 99% accuracy. This breakthrough is critical for the AI/ML community as matrix multiplication is fundamental to machine learning models, including large language models. The research utilizes a novel software system based on "inverse design," allowing the automated creation of structures by specifying desired functionalities first, which is a departure from conventional engineering methods. The team faced challenges with encoding negative coefficients due to heat conduction laws but successfully addressed this by optimizing separate structures for positive matrix components and later combining outputs. While practical applications remain limited, particularly for large-scale deep learning, the capability to detect heat sources and manage temperature changes in microelectronics could transform how we monitor device health and performance without additional energy consumption. Future work aims to enhance these structures' abilities to conduct sequential operations and improve programmability, paving the way for more robust applications in AI and beyond.