Neuromorphic technologies for next-generation cognitive computing

With the end of Dennard scaling and the consequent slow-down of Moore's law, researchers are looking to exploit new device, circuit and architectural concepts to build future systems that are exponentially more capable than the systems of today. One promising approach is Neuromorphic computing based on non-Von Neumann architectures, that could achieve orders of magnitude performance improvements over conventional systems on certain neural network training tasks. In this paper we review our recent work on implementing such a neuromorphic system on arrays of Phase-Change-Memory (PCM) devices. We present both experimental and simulation results that explore the impact of device non-idealities such as non-linearity, asymmetry, drift and variability on neural network performance.We also highlight circuit design approaches to enable speedup vs. conventional systems.