System Level Power Reduction for YOLO2 Sub-modules for Object Detection of Future Autonomous Vehicles

AI-based object detection is one of the critical design components for the success of level 3 or level 4 autonomous vehicles. Currently, GPU based system design is a popular way to implement Convolutional Neural Networks(CNN) or Recurrent Neural Networks(RNN) for the system. The thing is that the feasibility of the GPU-based implementation for real cars is quite low because of the huge power consumption of GPU. In this paper, we claimed that a holistic approach from system level to circuit level is necessary for ultra-low power design. As an initial step for the approach, we proposed system-level power reduction techniques that can be applied to advanced CNN algorithms such as YOLO2. By applying proposed system level techniques such as loop unrolling, declare small function as inline, arguments passing, branching, and common expressions elimination, we demonstrated that the proposed low-power schemes can be reduced up to 86.95% for YOLO2 sub-modules compared with the original one using system-level power-estimation CAD tools (Simple Scalar and WATTCH [1]).