A grain size of a crystal grain in a crystalline semiconductor film obtained by a thermal crystallization method using a metallic element is reduced. Thus, the number of crystal grains in active regions of a device is made uniform. The thermal crystallization method using a metallic element is performed for a semiconductor film formed on an insulating film formed at a lower temperature than that at formation of the semiconductor film and that at crystallization of the semiconductor film. By thermal treatment in a step of crystallizing the semiconductor film, stress of the insulating film is applied to the semiconductor film, thus causing distortion in the semiconductor film. When the distortion is caused, surface energy and a chemical potential of the semiconductor film are changed to promote the generation of a natural nucleus. Therefore, since a generation density of the crystal nucleus is increased, a grain size of a crystal grain can be reduced.