I spent a little bit of time reading this paper and thought I would share some preliminary results. As a brief summary, the author presents an approach to perturb the surface normal of a piece of geometry using a supplied heightmap and shows it to be equivalent to the standard approach. Practically speaking it produces results that are a little worse as it relies on the screen-space derivative functions (which are not implemented at full resolution on current hardware AFAIK). Now some images:

The image on the far right uses a method presented in the paper of using a gradient map computed from the heightfield. I don't think my implementation is quite correct for that yet... but it still produces pleasing results at a comparable speed to the normal mapping implementation pictured in the middle. The image on the left is the result of using the heightmap - the lower resolution of the screen-space derivative function should be apparent there.

Overall I am not entirely sure how useful it is, but it was a fun exercise. I'll try and add a few things to make my test app a little more user friendly and then post it up here for people to give a try.