Real-time gaze estimation has extensive applications in various fields, such as smart classrooms, advertising analysis, and smart homes. With the continuous maturation of neural network technology, gaze estimation under large models can achieve excellent accuracy and speed, but it requires high-computing processors, which results in overly large devices, thereby limiting the application scope of gaze estimation. Deploying a real-time gaze estimation system onto edge devices, albeit with a slight compromise in accuracy to ensure real-time performance, significantly enhances the practical value of estimation. This paper deeply integrates gaze estimation algorithms with FPGA by leveraging block-wise convolution and fusing single convolutions to address the limited on-chip memory of FPGA, thereby improving the parallelism of model inference. As a result, we achieved 32 frames per second on the ZYNQ7035 processor with an average power consumption of 6.4 watts.