I guess you’ve looked at Stefan Gustavsons simplex/perlin implementation already?!
Hm, I think OP is after some implementation that doesn’t use textures at all in the VS (not just an implementation that works in the VS). Although the optimizations mentioned in chapter 26 might be worthwhile looking at in any case.
I’ve never myself seen or heard about a simplex/perlin implementation not based on textures - it seems unavoidable for the pseudo-random numbers at the very least.
If you don’t consider VTF an option at all, then you might have to make do with precomputed noise attached as attributes to each vertex. Or?
However, VTF does work satisfactory on ATIs newer hardware (2600 series and later?) and obviously also on NV since the GF6800 series and later. And you would want to stay with NEAREST point filtering for the permutations anyway, so the restrictions on the texture formats for GPUs that are only SM3, just means you’ll be using a bit more memory for textures that are already small.