It's brutally expensive, unfortunately. Moving from 2D to 3D you go from needing O(n²) particles to O(n³) particles to fill a box. That's an awful increase in complexity, and it really limits the size of what you can do.
With that said, NVIDIA has a boatload of money, top-notch hardware skills and fluid simulation experts like Matthias Müller-Fischer working on the problem. NVIDIA Flex[1] is amazing to behold.
As a slight aside, is NVIDIA's motivation profit from the gaming industry or does this find industrial application outside of the obvious visual presentations?
Possibly, given that - besides all that recent neural-networks-on-GPU craze - people run e.g. CFD simulations on GPUs. I recall SpaceX in particular doing simulations for their rockets on NVIDIA hardware.
Flex, PhysX and Gameworks are all meant for games, but I would suspect that the game physics talent helped guide CUDA, which does have a lot of industrial applications.
They aquired Müller and PhysX when they bought Ageia in 2008. They changed Ageia's direction away from developing dedicated hardware (a Physics Processing Unit) to instead leverage the GPU for physics calculations.
