Non-volatile memory for neuromorphic AI accelerators
|
김상범 교수 서울대학교 The recent breakthroughs in deep learning have spurred interests in the development of novel computing architectures which can overcome downsides of conventional Von-Neumann computing architecture. Novel computing architectures utilizing non-volatile memory as synaptic devices have the potential to accelerate neural network computations, regarding both inference and training, in an area- and energy-efficient manner. To realize this potential, one needs to address various issues ranging from materials, devices, architectures, and algorithms. In this talk, the spiking neural network chip utilizing phase change memory cells as synaptic devices will be discussed focusing on how non-ideal characteristics of PCM cells as synaptic devices can be mitigated by material, device, circuit, and algorithmic level to enable scalable learning algorithm with spiking neural network. |