Radiofrequency Spintronic Neural Network Enabled by Electrically Modulated Magnetic Tunnel Junctions.

Abstract

Magnetic tunnel junctions (MTJs) in nanoscale have emerged as promising candidates for energy-efficient neuromorphic computing. As a pioneering demonstration, the radiofrequency (RF) neural network based on the intrinsic RF-to-DC conversion capability of MTJs features multilayer interconnectivity and native processing of RF inputs. However, most existing devices rely on magnetic field lines to modulate their behavior in neural networks, resulting in high energy consumption and increased area overhead. Moreover, the limited tunable bandwidth of the MTJs constrains the number of synapses per layer, thereby limiting the network's potential for scaling up. In this work, electrically tunable spintronic synapses and neurons based on three-terminal MTJs are experimentally realized, where the spin-orbit torque enables precise modulation of synaptic weight and neuron output frequency. The proposed methodoffers enhanced scalability and reduces energy consumption by a factor of 21. Furthermore, multilayer networks employing both fully connected and convolutional architectures, achieving 99.2% accuracy on drone classification and 92.0% on the Fashion-MNIST image dataset, are stimulated. The convolutional design notably reduces the number of required oscillator frequency channels. The results demonstrate the feasibility of scalable, high operational frequency, and energy-efficient all-spintronic neuromorphic systems, offering a compatible platform for future neuromorphic computing applications.