Hm, I wonder how they count the triangles rendered, since they could as well count the triangles which are rendered in shadow map passes, which would increase the amount of triangles rendererd per frame a lot. Like if your scene contains about 1000K tris, if you are rendering two shadow maps, the tris rendered could as well jump up to 1400K.

That being said, its completely possible to render a ton of triangles. Modern GPUs can process a load of vertexes in a very short time. The thing that matters is the pixel calculation complexity. Calculating lighting and several other effects are often done with textures (image space) nowadays which puts much more strain on pixel processing stage.