Skip to main content
If you don’t need to recompile code, you typically don’t need to install CUDA/cuDNN separately. Frameworks come with precompiled CUDA support, and the framework version corresponds to a specific CUDA version. Therefore, you only need to focus on the framework version and not the CUDA version independently.

​
Check Default CUDA/cuDNN Version

The CUDA version displayed by the nvidia-smi command only indicates the highest CUDA version supported by the driver, not the actual version of CUDA installed on the instance.
To check the default CUDA version installed in the platform’s built-in image (installation directory is /usr/local/):
bash
# Query the CUDA version in the platform's built-in image
ldconfig -p | grep cuda
    libnvrtc.so.11.0 (libc6,x86-64) => /usr/local/cuda-11.0/targets/x86_64-linux/lib/libnvrtc.so.11.0
    libnvrtc.so (libc6,x86-64) => /usr/local/cuda-11.0/targets/x86_64-linux/lib/libnvrtc.so
    libnvrtc-builtins.so.11.0 (libc6,x86-64) => /usr/local/cuda-11.0/targets/x86_64-linux/lib/libnvrtc-builtins.so.11.0
To check the cuDNN version installed in the platform’s built-in image:
bash
# Query the cuDNN version in the platform's built-in image
ldconfig -p | grep cudnn
    libcudnn_ops_train.so.8 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8
    libcudnn_ops_train.so (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so
    libcudnn_ops_infer.so.8 (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8
    libcudnn_ops_infer.so (libc6,x86-64) => /usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so
The version number is indicated by the number after .so in the output logs. If you have installed CUDA via conda, you can check it using the following commands:
bash
conda list | grep cudatoolkit
cudatoolkit               10.1.243             h6bb024c_0    defaults

conda list | grep cudnn
cudnn                     7.6.5                cuda10.1_0    defaults

​
Install Other Versions of CUDA/cuDNN

​
Method 1: Install using Conda

Advantages: Simple and easy to use. Disadvantages: Typically, header files are not included. If you need to compile code, you will need to install using Method 2.
bash
conda install cudatoolkit==xx.xx
conda install cudnn==xx.xx
bash
  conda search cudatoolkit
  Loading channels: done
  # Name                       Version           Build  Channel             
  cudatoolkit                      9.0      h13b8566_0  anaconda/pkgs/main  
  cudatoolkit                      9.2               0  anaconda/pkgs/main  
  cudatoolkit                 10.0.130               0  anaconda/pkgs/main  
  cudatoolkit                 10.1.168               0  anaconda/pkgs/main  
  cudatoolkit                 10.1.243      h6bb024c_0  anaconda/pkgs/main  
  cudatoolkit                  10.2.89      hfd86e86_0  anaconda/pkgs/main  
  cudatoolkit                  10.2.89      hfd86e86_1  anaconda/pkgs/main  
  cudatoolkit                 11.0.221      h6bb024c_0  anaconda/pkgs/main  
  cudatoolkit                   11.3.1      h2bc3f7f_2  anaconda/pkgs/main

​
Method 2: Install by Downloading and Installing the Package

CUDA Download Address: CUDA Toolkit Archive
Installation Method: After downloading the .run installation package:
bash
chmod +x xxx.run   # Add execution permission
./xxx.run          # Run the installation package
cuDNN Download Address: CUDA Deep Neural Network
Installation Method:
  1. Unzip the downloaded file.
  2. Move the dynamic libraries and header files to the corresponding directories:
bash
mv cuda/include/* /usr/local/cuda/include/
chmod +x cuda/lib64/* && mv cuda/lib64/* /usr/local/cuda/lib64/
  1. After installation, add the environment variable:
bash
echo "export LD_LIBRARY_PATH=/usr/local/cuda/lib64/:${LD_LIBRARY_PATH} \n" >> ~/.bashrc
source ~/.bashrc && ldconfig
The default image includes the most basic version of CUDA and cuDNN. If you have installed cudatoolkit via Conda, it will generally be used preferentially over the default installation.