A CNN Accelerator with Embedded Risc-V Controllers

Convolutional Neural Network is a promising technology in machine learning. Due to its vast computing and data requirements, it needs to be run with a specific accelerator to achieve reasonable energy efficiency. Improving the performance of accelerators has become the research hotspot. A mixed-precision structure can be used to improve hardware utilization, thus reducing area and power. However, the mixed-precision flow control is so complicated that it costs too much hardware resources. In this paper, a CNN accelerator with embedded RISC-V controllers is introduced to achieve flexible control at a very low cost. The ASIC synthesized results show that the proposed design area with two embedded cores is 5% less than the basic design.