The finite difference time domain (FDTD) method is one of the important methods for solving Maxwell’s equations in electromagnetics, and it is widely used. But it is time consuming when applied to the simulation of electrically large targets. In order to solve this problem, we take advantage of the parallel processing capacity of the graphics processor unit (GPU) together with the compute unified device architecture (CUDA). Taking a low pass filter as an example and using five million targeting grids, we realize three-dimensional FDTD high performance speed calculation with time-domain convolution perfectly matched layer (CPML) absorbing boundary. Experiments are carried out on the Quadro 4000 and Tesla M2050 GPUs with the Fermi architecture, whose error rate is within the range of  10-4, and can obtain 36 and 55 times faster speed than the CPU of the same period. The results show that the method has the characteristics of high precision, high efficiency, versatility and strong practicability.