The video memory backbuffer the screen will pan on will be something like 1024*1024. In most cases, he'll be able to scroll smoothly on a 640x480 screen; he's using a video memory blit from the background. But when the visible region he's panning on intersects the backbuffer border, he must then redraw the tiles onto the backbuffer. This redraw is slow but he may might be able to do it faster using a fairly large system memory bitmap and blitting portions of the syxtem memory bitmap onto the video memory bitmap. Occasionally he'd have to reach back on his tiles and blit them on to his system memory bitmap. But because this redraw would only rarely occurr, he could probably arrange for it on his map.
Having a larger surface than the screen eats memory unnecessarily. Video memory is still a precious resource and at most you will get a 2048 x 2048 map because that is the absolute maximum size of any texture on any card to date....I think.