Hi
When I put my old GTX 570 into my main rig to use for CUDA rendering in blender together with my new GTX 1060 6GB, the cards work independently if only one of them is selected, via CUDA_VISIBLE_DEVICES or otherwise, but when both cards are enabled in the same program it throws a invalid CUDA device error in blender. I can reproduce this using the simple multi-gpu sample from the CUDA 8 SDK, which will throw
"CUDA error at ../src/simpleMultiGPU.cu:121 code=46(cudaErrorDevicesUnavailable) "cudaStreamCreate(&plan[i].stream)" "
I have verified nothing else is running on the GPU, is this intended behavior when mixing a very old card with a new card, or a regression? If any of the GPUs are masked out via CUDA_VISIBLE_DEVICES, the program completes successfully
nvidia-bug-report.log.gz (129 KB)