I think the NVIDIA process is a special case.

In contrast, normal maps exported from zBrush have 3 dimensions where RGB corresponds to vectors in XYZ space. NVIDIA's "extrapolation", it would appear, would not give you the same detail as one would normally find. I cannot comment on Poser's implementation of normal maps but in my renderer a normal map and displacement mapping are equally fast to render which is dis-advantageous for normal maps since their silouettes aren't  perturbed.