This blog was originally published in the Linux on Power community, on May 31, 2017.
Some changes and issues have arisen since our first post about running GPUs in containers on POWER.
The most exciting news is that the CUDA images have been pushed to Docker Hub, so you no longer have to build your own cuda, cudann5, or cudann6 images for ppc64le. We’re working to get the others also pushed to Docker Hub, so stay tuned.
The nvidia-docker repo has undergone some recent changes. The biggest change is that image Dockerfiles are now hosted separately from the plugin code. For x86/amd images, see https://gitlab.com/nvidia/cuda. For POWER, see https://gitlab.com/nvidia/cuda-ppc64le. To build the nvidia-docker plugin, you still use the original github repository. Please see the post linked earlier for more on building that plugin.
The last update is that an issue has been discovered with the nvidia drivers for ppc64le. There was a kernel change that affected all platforms, but due to slight differences between code paths, only the NVIDIA drivers for ppc64le are affected. For more info, you can read this github issue: https://github.com/NVIDIA/nvidia-docker/issues/350#issuecomment-295438792. If you need a workaround, the issue in the kernel is fixed in upstream 4.10. We are looking into getting this fix backported into Xenial.
Again, many thanks to the folks at NVIDIA for their cooperation and support.