http://www.nvidia.com/content/quadro/ma ... gue_HR.pdfIn short, you get one GPU for OpenCL and one for OpenGL. This requires you to transfer from the OpenCL card back to system memory for display, so you have to disable OpenCL caching, but that's disabled by default anyway. The two things it does is that it allows you to take however long you want on OpenCL tools without locking up your UI, which is great if you have REALLY slow OpenCL tools. The other thing is that you can split the memory needed, and you do less task switching. So that's nice. The downside is, it's only noticeably faster when you have blocking between OpenCL and OpenGL, which only happens when you do simultaneous branching AND there's no CPU bottlenecks in between. If you do a LOT of really heavy OpenGL and OpenCL in the same comp, it's great, though.
Newer versions of OpenCL make it easier to have multiple devices running the same kernels, so it would be nice if Fusion supported multiple GPU's for OpenCL. Also, the Renderer3D's OpenGL renderer already breaks the render up into tiles and passes (and views, if you do stereo or depth of field). So there's a LOT of ways that the OpenGL rendering could be sped up with multiple GPU's. You can't, however easily speed up the 3D Viewers with multiple GPU's without requiring Quadros. If BMD wanted to sell a $9995 version of Fusion, multiple GPU support would be a great feature to add.