Energy-Efficient Brain-Inspired Computing by Electrostatic Gating of Disordered Networks

Dopant Network Processing Units (DNPUs)¹ are nanoscale electronic devices that could be promising building blocks for future brain-inspired computing systems. They consist of eight bias electrodes that can give rise to complex functionality to the network hopping sites they are deposited on top of. Here, we designed and characterized a more energy-efficient version of this non-linear classifier. Four of the electrodes are positioned on top of a 20 nm layer of aluminum oxide that separates them from a two-dimensional network of defects. These electrodes no longer draw a current, while they still electrostatically influence the electrostatic potential landscape of the device. The experiments show that these electrostatic gates non-trivially contribute to the functionality of the device. The device is able to solve all Boolean gates, while exhibiting a 9-30 times lower power consumption than reported originally by Chen et al¹, depending on the mode of operation. A surrogate model was trained to simulate the performance of an ensemble of these devices. MNIST was solved with a 95.7% accuracy and a 50 times lower power consumption than reported previously. Electrostatic gates prove to be an effective, yet energy efficient way to control devices based on disordered networks.