The invention provides a GPU-based convolution optimum matching layer boundary condition implementation method. According to the implementation method, by setting an auxiliary variable of a zero value in a simulation area and converting a calculation formula of the simulation area into the form identical to a formula of an optimum matching layer, calculation of the simulation area and an absorption layer can be completed in the same GPU task instead of being divided into multiple GPU tasks to be completed in a serial mode, and therefore calculation efficiency is improved; a compressed storage mode is adopted for the auxiliary variable, and access to the auxiliary variable is realized through a mapping array, so that the memory use amount of a GPU is effectively controlled, and the calculation ability of the GPU is maintained. Compared with existing implementation methods, the calculation efficiency is greatly improved on the condition that only the requirement for a minute quantity of memories is added, and therefore practical production capacity is improved.