Abstract: Back Projection (BP) is a classical time-domain imaging algorithm for processing synthetic aperture radar (SAR) signal, which can be used for fast factorized BP Algorithm to adapt to SAR sub-aperture imaging with long synthetic aperture time, large width and curved trajectory of the satellite. BP is used to image SAR sub aperture echoes in improved FFBP algorithm. However, its huge amount of computation makes the timeliness difficult to be satisfied in project. In this paper, the graphics processor unit (GPU) is used to solve this problem, as a coprocessor for the CPU. The paper proposes a solution around in CUDA-based to achieve FFBP algorithm for GEO SAR sub-aperture imaging. First, stream technology is used to hide the time delay caused by echo data transmission between blocks. At the same time, the scheme also avoids the process being switched frequently. Then, GPU large-scale concurrency of FFBP algorithm sub-aperture imaging is achieved by designing ultrafine granularity threads. By verified, the execution efficiency of the device is greater than 90% when use the CUDA solution to complete the FFBP algorithm sub-aperture imaging, its processing efficiency is 120 times better than CPU 32 threads concurrent program.