Model scale has very little to do with performance. I would recommend modelling at a real-world scale, that way it's a lot easier to figure out relative scale of different objects.
What will matter is how many objects, materials and vertices etc. are visible and processed each frame. For a topdown game the camera angle and camera distance will be very important in deciding how detailed your models should be.
If you know you want a distant camera, with a large view area, you probably want to keep model details to a minimum. On the other hand, If you want a close-up perspective, with few object visible at a time, then feel free to crank up the details.