Parallax Occlusion Mapping...

[sgi]

I've tested on other hardware/driver pairs and they all failed the same:
GeForce GTX460, 258.96 - FAILED
GeForce GTX460, 260.63beta - FAILED
I'll try to modify shader to find the source of troubles.
My first guess is that dynamic cycle count can cause that, GPU just cant calculate derivatives for tex2D sampling if number of cycles is different in the same quad of pixels - this also explains why lags depend on current view. Of-course driver can handle this scenario and invoke the same number of iterations on all 4 threads of the quad, but I think this is the place in driver where the problem can be.