Energy Efficient Fixed-point Inference System of Convolutional Neural Network

In this paper, we implemented a high accuracy, high throughput, low bit-width, and low power consumption convolutional neural network (CNN) inference system for handwritten digit recognition onto an FPGA. With reference to the 32-bit floating-point baseline, our 8-bit fixed-point design has zero accuracy loss and maintains 99.55% accuracy. Moreover, our improved 4-bit fixed-point design achieves 99.28% accuracy, which has less than 0.3% accuracy loss. Also, our 4-bit fixed-point LeNet-5 inference accelerator is implemented on FPGA Evaluation Kit and achieves a throughput of 2296 GOP/s with an energy efficiency of 213 GOP/s/W.