So I got a chance to configure nvidia-docker on top of ubuntu 16.04 and try some ffmpeg cuda based video encoding the GTX 1060. Everything runs ok but it looks like only two containers can access the GPU. This seems to be a limitation. Trying to launch a third container generates an error stating “No GPU found”. Looks like a hardware / software (sdk) limitation?
This is too bad. We do lots of video encoding and it’d be really nice to consolidate the 30+ boxes into one or two Docker systems that can handle 15+ encoding instances each :P
Just wanted to include this info for anyone else who’s trying to test ffmpeg and cuda on top of Docker !! Thanks again for the original post and insight !
References to things used:
NVIDIA/nvidia-docker
nvidia-docker - Build and run Docker containers leveraging NVIDIA GPUs
github.com
FFmpeg with Cuda support:
https://github.com/nightseas/gpgpu_applications/tree/master/docker/ffmpeg
