The soft-sensing model for solidification structure of continuous casting slab is complicated in algorithm, large amount of calculation and time-consuming in solution. The method based on the central processing unit (CPU) is difficult to meet the prediction needs of large-size casting. To improve the calculation efficiency, a cellular automaton ( CA) soft-sensing model based on graphic processing unit (GPU) heterogeneous parallelism is proposed. Firstly, the heterogeneous parallel algorithm of GPU-CA is designed to eliminate the data dependence and data competition among cells, which optimizes the parallelism degree among data. Secondly, a multi-stream task scheduling scheme is proposed to solve the problem of independent tasks waiting each other in single-stream, and improving the degree of task parallelism. Finally, two kinds of the steel produced by a large-scale continuous caster in a certain steel plant are used to test the model. The predicted results are in good agreement with the field experiment data, where equiaxed grain rate errors are about 1% and 1. 5% , respectively. The maximum relative error between temperature and measured temperature is 1. 37% . In the case of the same calculation accuracy as CPU, the speedup of GPU is hundreds of times, which greatly improves the computing speed of the model.