What’s the different between the GPU2 (second generation) and the GPU3 (third generation) client?

The differences from the donor perspective are relatively minimal. However, the underlying scientific differences are fairly dramatic. GPU3 is more stable, more efficient, and has a larger set of scientific capabilities. We now use OpenMM, our open-source molecular simulation toolkit. We also utilize OpenCL for ATI GPUs and CUDA for Nvidia.

Please see these two blog posts for more information: Prepping for GPU3 release and Open Beta Release of GPU3.