The physx gpu runtime had several CUDA branches, that didn't always evolve linearly (I tried to guess the SDK version from the minimum supported driver version when unspecified, though it may be not every dll in a given release is always recompiled):
- CUDA 2.0: in the beginning there was just PhysX 2.7.3, which eventually expanded by mid-2009 with every version between 2.7.1 and 2.8.1 being accelerated. 2.7.1, 2.7.4 and 2.7.6 were somehow dropped after some month (CM2 allegedly relies on this one version for some reason)
- CUDA 3.0: everything is back in, except for 2.7.2 (in a "supported set" that is already reminiscent of the present).
- CUDA 4~4.2: from just the release notes you'd think 2012 is when this happened, but judging from the dlls this had already happened in 9.11 (after more than a year of radio silence nonetheless, and after updating their installer toolset)
- CUDA 5(?): nothing much to add, except this was around the time they decided not to ship anymore the non-accelerated versions
- CUDA 6~6.5: like above, you would think that it's only with the big Tesla bye bye that something significant happened, yet dlls were already there in the previous release
Other funny juicy facts are that once upon a time somebody got Curie to reportedly flag physx support... and a guy was in touch for real with nvidia for physx to work on ATI cards (fat chance yeah, even today the best hope is ZLUDA)