In order to save some texture memory in a multi-step process, I am thinking of binding an FBO in order to update a small part of it with the scissor test enabled.
As I am not performing a glClear() in that case, what is the expected penalty for loading the data into on-chip memory ? Will the GPU restrict the data fetch to only those tiles intersecting the scissor box (it is very small compared to the overall texture, less than 5%) ?