I've seen OpenCL beta software running on windows. A guy from EA wrote a cloth simulation in OpenCL and was able to simulate and render hundreds of cloth shirts at 60hz on an Nvidia 8800GT. He also showed it on a 4xxx ATI and on a Quad Core CPU.
Also, CUDA and OpenCL work very similarly, so any case where someone wrote something in CUDA and got a certain speedup, you could get the same speedup by porting the code to OpenCL. CUDA is just Nvidia proprietary stuff.
The Bullet Physics engine is working on an OpenCL backend right now. They have a CUDA backend already. In early test with the CUDA backend, they said
Look Here
So they got a 100% improvement in performance just by offloading a part of the physics pipeline to the GPU. They have constraint solving working in a demo now, but I didn't see any benchmarks on that yet.