Probabilistic neuromorphic system using binary phase-change memory (PCM) synapses: Detailed power consumption analysis
Abstract
In this paper we investigate the use of phase-change memory (PCM) devices as binary probabilistic synapses in a neuromorphic computing system for complex visual pattern extraction. Different PCM programming schemes for architectures with- or without-selector devices are provided. The functionality of the system is tested through large-scale neural network simulations. The system-level simulations show that such a system can solve a complex real-life video processing problem (vehicle counting) with high recognition rate (>94%) and low power consumption. The impact of the resistance window on the power consumption of the system is also studied. Results show that the learning-mode power consumption can be dramatically reduced if the RESET state of the PCM devices is tuned to a relatively low resistance. Read-mode power consumption, on the other hand, can be minimized by increasing the resistance values for both SET and RESET states of the PCM devices.