GPU rigid body pipeline, 100k convex bodies in realtime, open source

Accelerated Computing CUDA CUDA Programming and Performance
1

You can simulate 100k rigid bodies on the GPU, using OpenCL. It works with recent drivers on Fermi or Kepler (and Radeons):

video: OpenCL gpu rigidbody pipeline, 64.000 boxes and convex barrels on a Tahiti at 25 frames/sec - YouTube

source and binary: https://github.com/erwincoumans/experiments/downloads