Efficient Neuromorphic Systems and Emerging Technologies: Prospects and Perspectives

Abstract

Recent advances in machine learning, notably deep learning, have resulted in unprecedented success in a wide variety of recognition tasks including vision, speech, and natural language processing. However, implementation of such neural algorithms in conventional “von-Neumann” architectures involve orders of magnitude more area and power consumption than that involved in the biological brain. This is mainly attributed to the inherent mismatch between the computational units—neurons and synapses in such models and the underlying CMOS transistors. In addition, these algorithms, being highly memory-intensive, suffer from memory bandwidth limitations due to significant amount of data transfer between the memory and computing units. Recent experiments in spintronics have opened up the possibility of implementing such computing kernels by single device structures that can be arranged in crossbar architectures resulting in a compact and energy-efficient “in-memory computing” platform. In this chapter, we will review spintronic device structures consisting of single-domain/domain-wall motion based devices for mimicking neuronal and synaptic units. System-level simulations indicate ∼ 100× improvement in energy consumption for such spintronic implementations over a corresponding CMOS implementation across different computing workloads.