Abstract
Deep neural networks (DNNs) computing has been rapidly moved from cloud to edge devices in many domains such as object recognition, natural language processing and video/image quality enhancement, etc. Neural network processing unit (NPU) leads a pervasive trend to accelerating neural network computation on edge devices. A NPU is a processor that is specialized for DNN algorithm computation. [4][5] Recently, a DNN based real-time video super resolution method which provides higher resolution video experience such as UHD from lower resolution input video, SD/HD/FHD etc.[2] However, it is difficult to apply DNN based super resolution because the algorithm consume high external memory bandwidth caused by huge input, intermediate and output data. In this paper, we describe our NPU architecture, called S-PIE (Samsung Perceptual Inference Engine) for super resolution application, and report our experience so far, including various components used for a verification framework of S-PIE.
Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
