I'd like some clarification on this too. In an ideal multithreaded context with every single polygon being handed out to a different GPU core the draw call completing with the last polygon is rasterised it'd take as long as the biggest polygon does to draw. However, I doubt polygons are "truly" multithreaded like this…