Although the existing granular-ball rough set (GBRS) model can use equivalence classes to represent knowledge and can process continuous and discrete data at the same time, it fails to make full use of the information of the granular-ball split before superimposing attributes when performing feature selection, and all boundary sample points still need to be split again. This paper proposes to make full use of the information of the boundary granular-ball before superimposing attributes, proposes an accelerated feature selection method based on GBRS, and gives a criterion to further guide the determination of the minimum number of samples in the granular-ball. Experimental results on benchmark datasets show that the learning accuracy and efficiency of the accelerated method are significantly improved compared with the GBRS algorithm.