In this work, we will focus on the use of phase change memory (PCM) to emulate synaptic behavior in emerging neuromorphic system-architectures. In particular, we will show that the performance and energy-efficiency of large scale neuromorphic systems can be improved by engineering individual PCM devices used as synapses. This is obtained by adding a thin HfO2 interface layer to standard GST PCM devices, allowing for the lowering of the Set/Reset currents and the increase of the number of intermediate resistance states (or synaptic weights) in the synaptic potentiation characteristics. The experimentally obtained potentiation characteristics of such PCM devices are used to simulate a 2-layer ultra-dense spiking neural network (SNN) and to perform a complex visual pattern extraction from a test case based on real world data (i.e. cars passing on a 6-lane freeway). The total power dissipated in the learning mode, for the pattern extraction experiment is estimated to be as low as 60μW. Average detection rate of cars is found to be greater than 90%.
Read full abstract