There are a few mechanics which are interesting to know.
1) The GPU and the drawing library we (currently) use (OpenGL, OpenGLES and DirectX) will always do the calculations for scaling, even if its 1280x720 -> 1280x720, it will be scaling by 1.0. Hence scaling is considered for free.
2) The GPU will always display an image at the given size you want it in, so the amount it needs to write is the same (this is limited by fillrate).
So if you look at 1 and 2 if you use a 8px texture or 1080p one it will take the same resources. So what you gain from having a 8px border texture is that you need to upload less to GPU, you need to decode less data. This means that it will load and display quicker, giving an illusion of faster operation.
There is another important limitation and that is that the GPU (atleast on embedded) needs to move the textures it reads into a special type of memory, if you take a bigger image than fits in that cache you need to fetch it multiple times and as such smaller textures are much much quicker to display, even if you need to scale them.
Overall, if you are contemplating going 8px texture or pre-scale it in photoshop, always let the GPU do it instead, its doing it for free and at times its even faster (since it means free up memory and load).
There are a lot of other optimizations you can do which is limit the amount of blended images you will use by pre-merging them, this is something which is not an issue on desktop however where the GPUs are so quick, on embedded however its huge. You should ofcourse not feel obligated to do this while developing, but if you are looking for optimizing something which you feel "finished" with, this can make a big difference.
I wrote a bit about this during my google summer of code (much of the info there is beagleboard and gsoc specific so some of it can be skipped)
http://elinux.org/BeagleBoard/GSoC/2010_...umentation