Highly Efficient Neuromorphic Computing Systems With Emerging Nonvolatile Memories

Increased interest in artificial intelligence coupled with a surge in nonvolatile memory research and the inevitable hitting of the "memory wall" in von Neuman computing(1) has set the stage for a new flavor of computing systems to flourish: neuromorphic computing systems. These systems are modelled after the brain in hopes of achieving a comparable level of efficiency in terms of speed, power, performance, and size. As it becomes more apparent that digital implementations of neuromorphic systems are far from approaching the brain's level of efficiency, we look to nonvolatile memories for answers. In this paper, we will build up highly-efficient neuromorphic systems by first describing the nonvolatile memory technologies that make them work, exploring methodologies for overcoming statistical device faults, and examining several successful neuromorphic architectures.