Radionuclides have been widely used in the fields of nuclear medicine, nuclear security and non-destructive testing, and their accurate identification is the basis of qualitative detection of radionuclides. In the portable nuclide recognition instrument, the traditional energy spectrum analysis method has the shortcomings of high delay and low recognition rate. This paper proposes a lightweight neural network model for nuclide recognition based on kernel pulse peak sequence and its FPGA hardware acceleration method. A lightweight and efficient neural network model is constructed by introducing depth-separable convolution and reciprocal residual modules, and using global average pooling to replace the traditional fully connected layer. For the network training data set, NaI (Tl) detector model was constructed through Monte Carlo toolkit Geant4 to obtain the analog energy spectrum, and then a simulator generated nuclear pulse signal sequences according to the energy spectrum, and 16 kinds of nuclear pulse signal data were constructed. Finally, the trained model is deployed to PYNQ-Z2 heterogeneous chip through optimization methods such as quantization, fusion and parallel computing to achieve acceleration. Experimental results show that the recognition accuracy of the proposed model can reach 98.3%, which is 13.2% higher than that of the traditional convolutional neural network model, and the number of parameters is only 2 128. After FPGA optimization and acceleration, the single recognition time is 0.273 ms, and the power consumption is 1.94 W.