Berkeley Lab Researchers Advance Thermodynamic Computing with New Design Framework

Researchers at Lawrence Berkeley National Laboratory have proposed a novel design and training framework for thermodynamic computing, potentially revolutionizing energy efficiency in machine learning. Published in Nature Communications, the study addresses key limitations of existing thermodynamic systems, which traditionally rely on reaching thermodynamic equilibrium and are restricted to linear algebra problems. The new approach, developed by Stephen Whitelam and Corneel Casert, utilizes nonlinear components to enable computations without waiting for equilibrium, effectively mimicking neural networks. Unlike classical and quantum computers that suppress thermal noise, thermodynamic computing harnesses these fluctuations as a power source, allowing for room-temperature operation and significantly reduced energy consumption. This breakthrough expands the applicability of thermodynamic computing to complex, nonlinear tasks typical of modern AI workloads. By demonstrating that nonlinear thermodynamic circuits can function like neurons, the team offers a promising path toward Beyond-Moore’s-Law microelectronics. This advancement could drastically lower the high energy costs associated with current digital computing methods, such as those used in large-scale search engines, marking a significant step in the development of low-power, energy-aware computing technologies.