An 8.1-to-353 TOPS/W Energy-Aware Deep-Learning Accelerator Supporting Dynamic Neural Networks

The demand for edge Al is rapidly increasing in various applications, including image classification, object detection, image segmentation, and key-point detection. Conventionally, deep-learning accelerators are designed for a fixed operating scenario, targeting a high accuracy with affordable energy consumption. However, the operating scenarios and performance requirements may vary. There are cases where low-energy operation is the first priority and a slight accuracy loss is tolerated. Therefore, an energy-aware solution supporting a variety of operating scenarios is required.