NVP: A Flexible and Efficient Processor Architecture for Accelerating Diverse Computer Vision Tasks including DNN

Compared with the CPUs and GPUs, the AI accelerators are able to achieve higher performance and energy efficiency for accelerating the DNNs. However, besides the DNNs, the computer vision also involves other tasks such as conventional image filtering and stereo matching. These tasks are not supported by the AI accelerators. In addition, the newly proposed DNN structures are not supported by the existing AI accelerators, making them difficult to catch up with the ever-evolving AI algorithms. To address this challenge, the Google has proposed the Pixel Visual Core (PVC) processor with a flexible architecture to accelerate diverse computer vision tasks including the DNNs while achieving higher efficiency. However, the architecture of the PVC is not well optimized, leading to limited energy efficiency. In this brief, we have proposed a flexible and efficient processor architecture (named NVP) with several design techniques to address the limitations of the PVC. The NVP is able to accelerate diverse computer vision tasks including DNN structures, conventional image filtering and stereo matching, while achieving significantly improved energy efficiency than the PVC and comparable energy efficiency with the AI accelerators.