Knowledgebase

Introduction

Compute Unified Device Architecture (CUDA) is a platform designed to perform parallel computing tasks using NVIDIA GPUs. Machine Learning programs use the GPU to parallelize and speed up tensor operations. Hence, the NVIDIA CUDA Toolkit accelerates the development and use of modern ML/AI applications such as Stable Diffusion and Large Language Models (LLMs).

CUDA consists of two APIs tools, the CUDA driver API, and the CUDA runtime API. The runtime API allows applications to plug into the CUDA platform and use GPU computing features in their programs. Upon deployment, the GPU drivers and driver API are available on a Rcs Cloud GPU server. Then, install the CUDA toolkit to enable the CUDA runtime API used by developers to take advantage of the GPU for parallel computations.

This guide explains how to install the NVIDIA CUDA Toolkit on a Ubuntu 22.04 server.

Prerequisites

Before you begin:

• Deploy a Ubuntu 22.04 Cloud GPU Server on Rcs with at least:

  • 1/12 NVIDIA GPU
  • 180GB Storage

• Using SSH, access the server

• Create a non-root sudo user account

• Switch to the new user account

    # su pythonuser

  > This guide uses the example user pythonuser, replace the account with your actual username

Install the NVIDIA CUDA ToolKit

To install the CUDA toolkit, you can use any of the following methods:

• Native Installation using a script or release file
• Using Conda with preinstalled Anaconda or Miniconda

Depending on your desired installation type, install the NVIDIA CUDA Toolkit on your server as described below.

Install CUDA Toolkit Natively (Recommended)

You can natively install CUDA toolkit using the distribution-dependent release file or an installation script. To avoid overwriting NVIDIA driver files, install CUDA toolkit using the latest installation script as described below.

1. Download the latest CUDA Toolkit run file

    $ wget https://developer.download.nvidia.com/compute/cuda/12.0.1/local_installers/cuda_12.0.1_525.85.12_linux.run

   The above command downloads the CUDA Toolkit version 12.0.1. To download the latest version, visit the CUDA Toolkit Archive files page.

2. Execute the installation .run file

    $ sudo sh cuda_12.0.1_525.85.12_linux.run

   * When prompted, read through the End User License Agreement (EULA), enter accept and press Enter to access the installation options.

   • On the Installation options list, press Space to deselect Drivers as they are already available on the Rcs Cloud GPU servers.
   • Scroll to the Install option and press Enter to start the installation process

   When the installation is successful, your output should look like the one below:

    ===========
    = Summary =
    ===========
   
    Driver:   Not Selected
    Toolkit:  Installed in /usr/local/cuda-12.0/
   
    Please make sure that
     -   PATH includes /usr/local/cuda-12.0/bin
     -   LD_LIBRARY_PATH includes /usr/local/cuda-12.0/lib64, or, add /usr/local/cuda-12.0/lib64 to /etc/ld.so.conf and run ldconfig as root
   
    To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-12.0/bin
    ***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 525.00 is required for CUDA 12.0 functionality to work.
    To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
        sudo <CudaInstaller>.run --silent --driver
   
    Logfile is /var/log/cuda-installer.log

3. To configure your server to work with the CUDA toolkit, perform the following operations:

   • Add the CUDA path to the system PATH

       $ echo "export PATH=/usr/local/cuda-12.0/bin${PATH:+:${PATH}}" >> /home/pythonuser/.bashrc

   • Add the CUDA Toolkit library path to the LD_LIBRARY_PATH

       $ echo "export LD_LIBRARY_PATH=/usr/local/cuda-12.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}" >> /home/pythonuser/.bashrc

   The above LD_LIBRARY_PATH command updates the CUDA toolkit link loader with the location of shared libraries

4. Activate the environment variable changes

    $ source /home/pythonuser/.bashrc

Install CUDA Toolkit using Conda

To install CUDA toolkit using Conda, verify you have either Anaconda or Miniconda installed on the server. Then, find the latest version on the Conda NVIDIA channel to install it on your server as described in the steps below.

1. Enter your Conda virtual environment, for example, env1

    $ conda activate env1

2. Install the CUDA toolkit

    $ conda install -c "nvidia/label/cuda-12.0.1" cuda -y

   The above command installs the cuda package version 12.0.1 that includes the CUDA Toolkit.

3. When the installation is successful, verify that the CUDA Toolkit is available in your local Conda packages

    $ conda list | grep "cuda-toolkit"

   Output:

    cuda-toolkit              12.0.1                        0    nvidia/label/cuda-12.0.1

4. To activate the CUDA toolkit on your server, create the Conda activate.d directory

    $ mkdir -p $CONDA_PREFIX/etc/conda/activate.d

   Conda uses the file env_vars.sh in the activation directory activate.d to load environment variables

5. Append the LD_LIBRARY_PATH to the env_vars.sh file

    $ echo "export LD_LIBRARY_PATH=$CONDA_PREFIX/lib${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}" >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh

6. Update the CUDA_PATH variable with the path to your installed CUDA packages and append it to the env_vars.sh file

    $ echo "export CUDA_PATH=$CONDA_PREFIX${CUDA_PATH:+:${CUDA_PATH}}" >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh

   The CUDA_PATH variable declaration allows NVIDIA sample programs to test CUDA installations using the toolkit executable files

7. Close your SSH session and re-access your server with the env1 Conda environment to apply the new system changes

    $ exit

Verify the Installation

1. Verify that the nvidia-smi drivers are working correctly

    $ nvidia-smi

   Output:

    +-----------------------------------------------------------------------------+
    | NVIDIA-SMI 525.125.06   Driver Version: 525.125.06   CUDA Version: 12.0     |
    |-------------------------------+----------------------+----------------------+
    | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
    | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
    |                               |                      |               MIG M. |
    |===============================+======================+======================|
    |   0  NVIDIA A40-1Q       On   | 00000000:06:00.0 Off |                    0 |
    | N/A   N/A    P8    N/A /  N/A |      0MiB /  1024MiB |      0%      Default |
    |                               |                      |             Disabled |
    +-------------------------------+----------------------+----------------------+

   If the command returns an error, the default NVIDIA drivers are overwritten, reinstall the drivers to integrate the GPU on your server again.

2. Verify that the NVIDIA CUDA Compiler (NVCC) package installed with the CUDA toolkit is available on your server

    $ nvcc --version

   Output:

    nvcc: NVIDIA (R) Cuda compiler driver 
    Copyright (c) 2005-2023 NVIDIA Corporation 
    Built on Tue_Jul_11_02:20:44_PDT_2023 
    Cuda compilation tools, release 12.0, V12.0.140
    Build cuda_12.0.r12.0/compiler.32267302_0

3. To verify that CUDA-based programs run on the server, clone the CUDA Samples Git repository that contains test programs

    $ git clone https://github.com/NVIDIA/cuda-samples.git

4. Switch to the deviceQuery sample program directory

    $ cd cuda-samples/Samples/1_Utilities/deviceQuery

5. Compile the sample program:

    $ make

6. Run the compiled file

    $ ./deviceQuery

   If the CUDA program runs successfully, your output should look like the one below:

    CUDA Device Query (Runtime API) version (CUDART static linking)
   
    Detected 1 CUDA Capable device(s)
   
    Device 0: "NVIDIA A40-1Q"
        CUDA Driver Version / Runtime Version          12.0 / 12.0
        CUDA Capability Major/Minor version number:    8.6
        .
        .
        .
        Result = PASS

   As displayed in the output result, the CUDA toolkit is now successfully installed and available on your server

Check System Compatibility for Other CUDA Toolkit Versions

Depending on your desired target CUDA Toolkit version you intend to install on your system, consider the following requirements to successfully run the toolkit.

• Supported NVIDIA Graphics Drivers
• Minimum Kernel version
• Requirements of your target CUDA-based applications

The CUDA version 12.0.1 you installed in this article runs on a server with the following specifications:

• GCC Version: 11.3.0
• NVIDIA graphics driver version: 525.125.06
• CUDA Driver API Version: 12.0
• Linux kernel version: 5.15.0-75

Depending on your target CUDA version, verify the system compatibility values as described below.

Verify the Attached System GPU

View the Linux system properties and search for the Nvidia graphics card

$ lspci | grep -i nvidia

Output:

6:00.0 VGA compatible controller: NVIDIA Corporation GA102GL [A40] (rev a1)

The above output displays a CUDA-compatible A40 GPU attached to the server.

View the GCC Version

The GNU Compiler Collection (GCC) is necessary when developing or compiling CUDA applications. Verify the installed GCC version on your system

$ gcc --version

Your output should look like the one below:

gcc (Ubuntu 11.3.0-1ubuntu1~22.04.1) 11.3.0 
Copyright (C) 2021 Free Software Foundation, Inc. 
This is free software; see the source for copying conditions.  
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Verify the Installed NVIDIA Drivers

Using the nvidia-smi command, verify if NVIDIA GPU drivers are available on your server

$ nvidia-smi

Output:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.125.06   Driver Version: 525.125.06   CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA A40-1Q       On   | 00000000:06:00.0 Off |                    0 |
| N/A   N/A    P8    N/A /  N/A |      0MiB /  1024MiB |      0%      Default |
|                               |                      |             Disabled |
+-------------------------------+----------------------+----------------------+

As displayed in the output, the server has the following installed driver versions:

• NVIDIA Accelerated Graphics Driver: 525.125.06
• CUDA Driver API version: 12.0

View the System Kernel Version

The CUDA Toolkit requires a recent version of the Linux kernel higher than the minimum required kernel version for a specific CUDA Toolkit version. Run the following command to view your available server Kernel version

$ uname -r

Output:

5.15.0-75-generic 

As displayed in the output, 5.15.0-75 is the available Linux kernel version compatible with the CUDA Toolkit version 12.0.1 installed on the server.

View the available Kernel headers

$ sudo apt list linux-headers-$(uname -r)

Output:

linux-headers-5.15.0-75-generic/jammy-updates,jammy-updates,jammy-security,jammy-security,now 5.15.0-75.82 amd64 [installed]

If the installed headers don't match the Kernel version, install the correct headers using the following command

$ sudo apt-get install linux-headers-$(uname -r)

Conclusion

In this guide, you installed the NVIDIA CUDA Toolkit on a Ubuntu 22.04 Cloud GPU server. You can develop and compile high performance GPU-powered applications on your server by integrating the toolkit with additional packages. For more information, visit the following official documentation resources.

• NVIDIA CUDA Installation Guide
• NVIDIA CUDA Compatibility
• CUDA Runtime API and Driver API Comparison