This is only true on discrete video cards that have local memory and lots of bandwidth. Even there, the fill rate may be big, but it is still finite.
On integrated chips (~70% of the market by volume and rising) and on embedded systems where the GPU and the CPU share the memory subsystem, blitting costs memory bandwidth, and the GPU can't do it any faster than the CPU.
Yep, exactly. Take nvidia Tegra2, for example. It's fine on small screens like phones, but attach a large screen (or two) to it for a tablet, and your memory bandwidth is the main limiting factor for doing redraws. If you clear the entire screen on every frame, you max out under 25fps -- and that's just to clear the screen, and doesn't give you any time to spend to render anything. Try and do anything interesting and you quickly drop below 10fps. So you end up having to do a bunch of tricks to limit what you redraw. Forget full-screen video with a high-res screen, though.
True, though even IGPUs have decent-sized texture caches which absorb some of the load, and embedded GPUs additionally usually have embedded frame buffer memory. In any case, blitting isn't anywhere near the crippling operation it once was.
This is only true on discrete video cards that have local memory and lots of bandwidth. Even there, the fill rate may be big, but it is still finite.
On integrated chips (~70% of the market by volume and rising) and on embedded systems where the GPU and the CPU share the memory subsystem, blitting costs memory bandwidth, and the GPU can't do it any faster than the CPU.