In the paper, how is the compression of the switch-transformer c2048 to 25.4x calculated?

hi, 
i am trying to understand qmoe， but it is too hard to figure it out that how you compute the theoretical compression limit of 25.4x of c2048 switch transformer. could you provide detailed computation process? 
