Parallel computation is an effective way to improve the efficiency of audio-frequency magnetotelluric (AMT) data inversion. In order to realize 2D AMT data inversion with high accuracy and high speed by using the powerful computing ability of GPU and parallel computing technology,the parallel algorithm of 2D AMT data inversion based on compute unified device architecture (CUDA) was studied. Firstly, the serial computation of 2D AMT data inversion was realized by using finite element method and adaptive regularization inversion algorithm, and then the parallel code based on CPU + GPU was written in the environment of PGI Visual FORTRAN + CUDA 5.5. The frequency cycle in forward modeling, the sensitivity matrix calculation in inversion and the regularized solution of inverse equations were paralleled. The validity and accuracy of the parallel algorithm were verified by forward and inverse simulation of theoretical models with different complexity. The comparison of numerical simulation results under different model scales and different modes shows that the parallel CUDA algorithm based on CPU + GPU is less time-consuming in sensitivity matrix calculation and inversion equation regularization than the traditional CPU serial algorithm, the maximum acceleration ratio can be more than 10 times. Finally, the parallel algorithm is applied to the 2D inversion of the measured AMT data in a mining area, and good results are obtained.