Yeah, as a game dev I would say the same. And PhysX runs on AMD computers, too, since it actually doesn’t even use the GPU. So PhysX can run on the CPU as well, and I guess Flex could do the same to some extend. But hell, I tried Flex, and some features did even kill my GTX 970!
Nvidia is currently promoting their Pascal chips for self-driving cars and some AI processing. It ain’t just a coincidence that GPUs are of more use than just graphics. This heavy paralleled processing can do more than pixels. What I’m saying is, that the current solution to make an API like PhysX, work on multiple platforms by offloading to the CPU, won’t work in future more advanced APIs! There is no way around hardware acceleration in the future it seems. And it’s not simple to make an API like FlameWorks, VXGI, or Flex either. So that one just could support both CUDA and OpenCL.
The problem is this. There is no alternative to Flex right now. I don’t know much about Havok, except that they got bought by Micro$oft from Intel. Though, actually Flex is still in Beta and not released yet really. Maybe by then Havok will come up with a unified particle based physics engine, too? But this would be just the physics, meaning the capabilities of FlameWorks and VXGI are still GPU locked. And only huge triple A studios could afford to develop their own versions of such APIs.