安全自主可控的国产化FPGA神经网络部署框架研究OACSTPCD

Aiming at the potential security issues arising from the existing deep learning edge applications relying on non-domestic FPGA architectures and encrypted IP implementations,which are difficult to quickly deploy on domestic FPGA platforms with insufficient IP and still under development,a neural network hardware deployment framework towards domestic FPGA is designed to achieve secure,autonomous and controllable deployment of domestic FPGA neural networks.In addition,the fast pipelined convolutional circuit and the im2col conversion circuit are designed to rapidly compute the systolic array.The experimental results demonstrate that the framework is capable of generating networks,such as Lenet5,with up to 147.8 GOPS for 16-bit data type and 0.024 ms running time,which repre-sents a 5.13x improvement in throughput and a 6.67x decrease in time,respectively.An evaluation of im2col operation on RTL design yields up to 124.9x gains over CPU on Intel Xeon E-2276M.