[tayistay]

Subdivision surfaces aren't so bad on the GPU. Since each limit vertex is a linear combination of control vertices, you can precompute the stencils to evaluate limit vertices. Those stencils can all be evaluated in parallel in a compute kernel. Generally, the base mesh topology isn't changing in an animation, so this works fairly well.

For Catmull-Clark, the regular parts of the mesh (quads with valence 4 vertices) are just b-splines and thus can be evaluated using the hardware tessellator.