I have a system with a NVIDIA 1080Ti, ubuntu 18.04 .2 LTS, linux driver version 396.54, Cuda 9.0 and cuDNN 7.4.1 installed with tensorflow 1.12
My code runs fine on the base machine. However, I wanted to try to bench the execution times of an identical docker container. unfortunately, the current nvidia docker image is:
I realized that the more recent nvidia drivers can be activated in the settings panel. It seems cuda 9.0 runs fine on them. Dunno tho if a container can have a more recent cuda version.
However, finding a container to match any particular CUDA, cuDNN, and TensorFlow version will be difficult because each container was released with the latest software available at that time. So, for example, all TF 1.12 containers are built using the CUDA 10.0 toolkit. It will likely be easier to choose a reference container and then setup your bare-metal installation to match that. For documentation about what component versions are used in each container release see the container release notes, https://docs.nvidia.com/deeplearning/dgx/tensorflow-release-notes/
Newer drivers are backward compatible with older CUDA releases. So if you upgrade only your driver, you will be able to run any NGC contaienr as well as your existing CUDA 9.0 tensorflow installation.