Highly pipelined Accelerator for Convolutional Neural Network

We propose a highly pipelined accelerator (HIPA) for the convolutional neural network (CNN). The CNN achieves remarkable performance in computer vision application with high computational power and low energy efficiency than conventional algorithms. Recent CNN accelerators with energy efficiency perform pre-loading in the internal memory by reducing data transmission from off-chip. We focus on the input data loading time to the processing unit for higher throughput. HIPA reduces loading time with optimized internal memory structure. For throughput, we propose a continuous snake-scan order to maximize data reuse. HIPA shows the input data loading time reduction by over 37% from VGGNet-D, and speed up to 530GOp/s.