A variability study of PCM and OxRAM technologies for use as synapses in neuromorphic systems

Overview & Synopsis

The human brain is made of a large number of interconnected neural networks which are composed of neurons and synapses. With a low power consumption of only few Watts, the human brain is able to perform computational tasks that are out of reach for today’s computers, which are based on the Von Neumann architecture. Neuromorphic hardware design, taking inspiration from the human brain, aims to implement the next generation, non-Von Neumann computing systems. In this thesis, emerging non-volatile memory devices, specifically Phase-Change Memory (PCM) and Oxide-based resistive memory (OxRAM) devices, are studied as artificial synapses in neuromorphic systems. The use of PCM devices as binary probabilistic synapses is studied for complex visual pattern extraction applications, evaluating the impact of the PCM programming conditions on the system-level power consumption.A programming strategy is proposed to mitigate the impact of PCM resistance drift. It is shown that, using scaled devices, it is possible to reduce the synaptic power consumption. The OxRAM resistance variability is evaluated experimentally through electrical characterization, gathering statistics on both single memory cells and at array level. A model that allows to reproduce OxRAM variability from low to high resistance state is developed. An OxRAM-based convolutional neural network architecture is then proposed on the basis of this experimental work. By implementing the computation of convolution directly in memory, the Von Neumann bottleneck is avoided. Robustness to OxRAM variability is demonstrated with complex visual pattern recognition tasks such as handwritten characters and traffic signs recognition.