Recently, the deep learning project needs to be deployed in the docker environment, so I can't avoid jumping into the pit of nvidia docker. First, I confirm that it is a very useful tool, but it is too clean and there is nothing in it. It takes some time to configure it. Here, the difference between the three image file versions of the unified cuda version is recorded as follows:
only from docker hub/nvidia/cuda
It can be seen that the image files of the three versions are different in size, and the specific differences are:
base version:
This version starts from cuda9.0 and includes the minimum (libcudart) for deploying pre-built cuda applications.
If users need to install the cuda package they need, they can choose to use this image version, but if they want to save trouble, it is not recommended to use this image, which will cause a lot of trouble.
runtime version:
This version extends the base image by adding all the shared libraries in the cuda toolkit. This image can be used if using a prebuilt application with multiple cuda libraries. But if you want to compile your own project with the help of the header files in cuda, you will get an error that the file cannot be found.
devel version:
Extend the running image by adding compiler toolchain, test tools, header files and static libraries, use this image to compile cuda applications from source code.