It is shown that the form of semiconductor yield models, and their predictions, are to a large extent affected by the size distribution of all defects and the spatial distribution of fatal defects. As the effects of these and other assumptions on yield models are rarely described in the literature, the author examines them and develops scaling rules for the average number of fatal defects per chip. To trace differences in the predicted yield to the various assumptions, the treatment compares a simple Poisson model with a compound Poisson model and shows that, when appropriate scaling rules are used, yield predictions of the simple Poisson model are accurate for a new product with chips with areas up to an order of magnitude larger than chips of existing products.>