降低实现神经网络的FPGA硬件资源消耗的方法研究

Research on Reduc FPGA for Implementing Neural Networks

为了满足训练神经网络的计算要求,将浮点数据替换为定点数据,使用更少的比特数,从而提高性能密度。虽然使用定点数据会降低精度,但只要性能比精度更关键,就可以使用定点数据。定点数据类型用于神经网络,可以降低使用FPGA实现神经网络的硬件资源消耗。

In order to meet the computing requirements for training neural networks,floating-point data is replaced with fixed-point data using fewer bits to improve performance density.Although using fixed-point data may reduce accuracy,it can be used as long as performance is more critical than precision.Fixed-point data types are used for neural networks and can reduce the hardware resource consumption of implementing neural networks on FPGA.