Nowadays, deep learning architectures like CNN have proven their superiority in image recognition tasks. To effectively deploy CNN networks in practice, especially for AIoT applications, it is essential to find a network model that offers good recognition performance with a small size and a limited number of parameters. Additionally, when deploying to hardware, to ensure rapid task execution, an FPGA-based approach is a suitable choice due to its parallel processing capabilities, low power consumption, low latency, and reconfigurability. In this paper, the authors propose a solution for searching and implementing an optimal CNN model on an FPGA chip for handwritten digit recognition using the MNIST dataset. The comparative results of recognizing 10,000 MNIST image samples on FPGA, CPU, and Raspberry Pi 4 microcontroller demonstrate that the proposed solution achieves high accuracy (98.78%) while maintaining low execution time (36 s). This indicates the potential application of the proposed solution in deploying deep learning tasks in practical applications.