nVidia has released a new BETA driver with CUDA 5. While the driver has caused a few problems with Collatz (e.g. returns results but they are all invalid), I wonder whether it would solve the multi-gpu issue with nVidia. Since the admins are aware and not doing much, the question is whether it is driver or app related.

For CUDA apps, one would just get the device parameter from the command line. But, because OpenCL lists all GPUs from all vendors but BOINC (due to the design) can only list devices by vendor, the BOINC developers had to come up with a day to select the correct OpenCL device when both AMD and nVidia are installed. For example, if you had both an nVidia and an AMD card installed, OpenCL considers them device 0 and device 1 whereas BOINC considers them to be device 0 for nVidia and device 0 for AMD. The issue with nVidia needing to be in slot 0 probably has to do with the order that BOINC numbers the devices when translating between OpenCL and CUDA/ATI. e.g. If BOINC sees them as both device 0, then when it runs the apps, one remains 0 while the other has to be changed to 1 in order for OpenCL to work properly.