Neuromorphic Computing with Resistive Switching Devices.

Abstract

Resistive switches, commonly referred to as resistive memory (RRAM) devices and modeled as memristors, are an emerging nanoscale technology that can revolutionize data storage and computing approaches. Enabled by the advancement of nanoscale semiconductor fabrication and detailed understanding of the physical and chemical processes occurring at the atomic scale, resistive switches offer high speed, low-power, and extremely dense nonvolatile data storage. Further, the analog capabilities of resistive switching devices enables neuromorphic computing approaches which can achieve massively parallel computation with a power and area budget that is orders of magnitude lower than today’s conventional, digital approaches. This dissertation presents the investigation of tungsten oxide based resistive switching devices for use in neuromorphic computing applications. Device structure, fabrication, and integration are described and physical models are developed to describe the behavior of the devices. These models are used to develop array-scale simulations in support of neuromorphic computing approaches. Several signal processing algorithms are adapted for acceleration using arrays of resistive switches. Both simulation and experimental results are reported. Finally, guiding principles and proposals for future work are discussed.