Deferred rendering is amazing to draw hundreds of realtime lights. But the CPU cost is quite high right now, because each light is one draw call. It would be a big optimization on low end platforms to render all the point lights in one draw call by leveraging GPU instancing.
If many small point lights are being rendered, very cool effects and atmospheres can be achieved. The GPU cost isn't high even on mobiles in this case, but the CPU cost prohibits that use case right now.