Tenfold power consumption reduction in reconfigurable nonlinear-processing units through capacitive control

Reconfigurable nonlinear-processing units (RNPUs) are silicon-based, multi-input-multi-output nanoelectronic devices that enable complex functionality and are promising building blocks for future brain-inspired computing hardware. However, in existing implementations, all electrodes draw current, resulting in substantial power dissipation. Here, we present a modified device design in which half of the electrodes are capacitively coupled to the silicon region through a dielectric layer, while the others are in direct electrical contact. This configuration allows electrostatic control of the potential landscape with minimal leakage current. We experimentally demonstrate all standard Boolean logic gates at room temperature consuming up to 10 times less power than reported for the original RNPU, which operated at 77K [Chen et al., Nature 577, 341 (2020)]. Simulations based on an RNPU surrogate model yield classification accuracies of 95.9% ± 0.2% for the MNIST handwritten digits dataset and 86.7% ± 0.2% for the FashionMNIST dataset with a similar reduction in RNPU power. These results demonstrate that capacitive electrodes provide an effective and energy-efficient means of controlling in-materia computing devices, opening a pathway for energy-efficient RNPU-based architectures.