OpenSubDiv demonstrates real-time subdivision
I don't mean to derail the conversation, just wanted to point out that the OpenSubdiv video linked is from last year and only shows uniform subdivision.
We have since released the "feature adaptive" version of the code which fully uses the dynamic tessellation capabilities of GPUs and scales much much better. Here is a video that was shot in Las Vegas last november, using Jesse Sandifer's "Turtle Barbarian" Mudbox sculpture.
We extracted the geometry from the original sculpture to Ptex displacement textures at a level of roughly 1/4 of a billion quad faces. Because we are executing the limit surface interpolation directly on the GPU, we only have to articulate the original control vertices of the coarse mesh (some 17000 of them), which allows us to animate the geometry in real time. The GPU tessellation also allows us to dynamically adapt the density of the geometry emitted each frame based on the position of the camera very smoothly, without the usual pops of mesh decimation...
I will probably have better visuals to link in the next few weeks...