It is indeed for generation on the CPU but i guess it still applies for the GPU.
Problematic with this approach is btw that you would need to implement a volumebased physicsengine or a physicapproach on the GPU that works with your extracted meshes etc.
Its hard to do, i think, if you need physics, that is.
But i didnt implement any of those algorithms on the GPU and i dont plan to, so i cant help that much with that. I just wanted to point you to some maybe helpful things