GF-4 is the first high spatial-resolution geostationary satellite of China. Surface reflectance products are of great value for evaluating ecological environment and reducing disasters. GF-4 atmospheric correction algorithm estimates surface reflectance, iteratively calculates the minimum residual error of observed and calculated apprent reflectance to get aerosol optical depth, and constructs a 6SV look-up table to obtain the surface reflectance results. Based on the accuracy verification of GF-4 surfacereflectance products, considering the high-efficiency requirements of data analysis and product production, the computationally complex steps are designed based on the graphics processing unit (GPU) to optimize the performance of threads and registers. The results show that the GPU-accelerated atmospheric correction algorithm has great advantages in performance and energy consumption. Compared with sequential execution, a scene of GF-4 PMS image data with 10240×10240 pixels has an overall speedup of 57.3, while the overall energy consumption is only 15.5% of sequential execution.