A team led by Giuseppe Romano at MIT’s Institute for Soldier Nanotechnologies developed an analog computing method that uses waste heat from electronic devices for data processing without relying on electricity. Instead of encoding inputs as binary 1s and 0s, the method represents data as a set of temperatures based on the heat already present in a device.
Heat moves through tiny silicon structures whose layout is designed by a physics-based optimization algorithm created by the researchers. The flow and distribution of that heat through those structures carry out the computation, and the output is represented by the power collected at the other end. The researchers used the structures to perform matrix vector multiplication, a core mathematical operation used by machine-learning models such as large language models to process information and generate predictions. The results were more than 99% accurate in many cases.
Significant obstacles remain before the approach can be scaled for modern deep-learning models. The researchers said there are challenges in tiling millions of these structures together. As the matrices become more complicated, results become less accurate, particularly when there is a large distance between the input and output terminals.
The method may have a nearer-term role in electronics by helping detect problematic heat sources and measure temperature changes without consuming extra energy. That could reduce the need for multiple temperature sensors that currently take up chip space. Caio Silva said the work reverses the usual approach to heat in electronics by treating it not as a waste product to remove, but as a form of information itself.
