Acceleration of Machine Learning with magnetic domain walls: Magnetic Neuromorphic Devices

See-Hun Yang, Ching-Tzu Chen
IBM Research - Almaden, California, United States

Keywords: artificial intelligence, deep machine learning, neurmorphic devices, magnetic domain wall devices, spin transfer torques

IBM proposes Deep Machine Learning using specialized hardware based on current-driven magnetic domain-wall motion. Magnetic domain-wall motion exhibits several highly desirable characteristics for in-data training of deep-learning neural network models. Although there is rapid progress in improving the performance of traditional chip architectures for neural network applications using GPUs or custom designed ASIC, these suffer from the von Neumann bottleneck which ties the compute unit to the memory. With ever increasing amount of data and complexity of the neural networks, the communication between compute unit and memory comes at the expense of compute efficiency, rendering it impossible to train neural networks on mobile devices, drones, and other edge devices. In comparison, our approach encodes the synaptic weights directly on the hardware level in the domain-wall positions, which enables updating the weights by electrical control and reading the weights by sampling the device conductance, eliminating the need for data shuttling. When fully optimized, the magnetic neuromorphic device array can potentially speed up training by ~4 orders of magnitude and reduce power consumption by ~2 orders of magnitude, thereby realizing Machine Learning on edge devices and dramatically shortening the training time for image classification, speech recognition, and language translation.