Abstract: In order to solve the problem of long calculation time and low efficiency when using dimensionless least square finite element method to calculate the two-dimensional transient flow field of transformer, GPU was proposed to accelerate the transient fluid field program in parallel. The two most computation-intensive parts of the transient fluid field calculation program, namely the formation of element stiffness matrix and the solution of sparse linear equations, were transplanted to GPU for calculation, thus greatly reducing the calculation time. Meanwhile, the cross-linked list method and CSR sparse storage structure were adopted to store the nonzero elements in the sparse matrix of the equations to reduce memory consumption. The square-cavity driver flow model was used to verify the effectiveness of GPU parallel program, and the speedup radio of parallel program increased with the size of cavity model scale. The GPU parallel program was applied to the transient fluid field simulation analysis of transformer winding model and the analysis results show that the acceleration ratio of GPU parallel program reaches about 16 times compared with serial program. The GPU-based parallel computing method implemented in this paper lays a foundation for the transient fluid field simulation of product-grade transformer.