What CUDA is and why you need to know your version

CUDA is a toolkit that lets your graphics card (GPU) do heavy computing work instead of your main processor. If you use machine learning software, video editing tools, or scientific programs, they often need a specific CUDA version to run properly. Checking your version tells you whether your system is set up correctly and whether you need to update.

Your CUDA version and your graphics card driver version are separate things — you can have one without the other working correctly. Knowing which version you have prevents the frustration of installing software that claims it needs CUDA 12 when you only have CUDA 11 installed.

Key Takeaways

  • On Windows, open Command Prompt and type nvidia-smi to see your CUDA version in seconds.
  • On Mac and Linux, the same nvidia-smi command works if CUDA is installed, or you can use nvcc --version to check the compiler.
  • If the command returns "not found" or "not recognized", CUDA is not installed on your system.
  • Your CUDA version appears in the top right corner of the nvidia-smi output — look for a number like 12.0 or 11.8.

Checking CUDA on Windows using Command Prompt

Open Command Prompt by pressing the Windows key, typing cmd, and pressing Enter. You should see a black window with a blinking cursor.

Type this exactly: nvidia-smi and press Enter. If CUDA is installed and your graphics card is recognized, you will see a table with information about your GPU. Look at the top right corner of the output — you will see a line that says something like "CUDA Version: 12.0" or "CUDA Version: 11.8". That number is your CUDA version.

If you see an error message saying "nvidia-smi is not recognized as an internal or external command", CUDA is not installed on your system, or the installer did not add it to your system path. You will need to install CUDA from NVIDIA's website or reinstall it.

Checking CUDA on Mac and Linux using the terminal

Open your terminal process. On Mac, press Command + Space, type terminal, and press Enter. On Linux, right-click the desktop or use your process menu to open a terminal.

Type nvidia-smi and press Enter. The output works the same way as Windows — your CUDA version appears in the top right. If you see "command not found", try the alternative command instead.

If nvidia-smi does not work, type nvcc --version and press Enter. This checks the CUDA compiler instead of the driver. You will see output that includes "release X.X" — that X.X is your CUDA version. This method works even if nvidia-smi fails, though it is less common on Mac.

Understanding what nvidia-smi actually shows you

The nvidia-smi output contains more than just your CUDA version. The table in the middle shows your graphics card name, how much memory it has, and how much memory is currently in use. Below that, you might see running processes using the GPU.

The number in the top right corner is what matters for checking compatibility. If a program says it needs CUDA 11.8 or higher and you have CUDA 12.0, you are fine — newer CUDA versions are usually backward compatible. If you have CUDA 11.4 and need 11.8, you will need to update.

Do not confuse the CUDA version with the driver version. nvidia-smi shows both — the driver version is listed separately, usually on the left side. You can have a new driver with an old CUDA version, or vice versa, depending on when each was last updated.

What to do if CUDA is not installed

If the nvidia-smi command returns an error on any operating system, CUDA is not installed. This does not mean your graphics card is broken — it means the CUDA toolkit has never been installed, or the installation failed.

Before installing CUDA, check that your graphics card is NVIDIA-based. AMD and Intel graphics cards do not use CUDA. You can see your card name in the nvidia-smi output if it is installed, or in your system settings if it is not. On Windows, right-click the desktop and look for NVIDIA Control Panel. On Mac and Linux, check your system information or graphics settings.

If you have an NVIDIA card and want to install CUDA, visit NVIDIA's CUDA Toolkit read page, select your operating system, and follow the installer. The process takes 10 to 20 minutes. After installation, restart your computer and run nvidia-smi again to confirm the version.

Checking CUDA version in Python or other programs

If you use Python with machine learning libraries like TensorFlow or PyTorch, you can check CUDA from inside Python itself. Open Python or a Jupyter notebook and type:

import torch (if you use PyTorch) or import tensorflow as tf (if you use TensorFlow). Then type torch.version.cuda or tf.sysconfig.get_build_info()['cuda_version']. The output shows which CUDA version that library thinks is installed. If this differs from what nvidia-smi shows, your library may not be using GPU acceleration correctly.

Frequently Asked Questions

Can I have multiple CUDA versions installed at the same time?

Yes. You can install CUDA 11 and CUDA 12 on the same computer, and different programs can use different versions. nvidia-smi shows the latest version by default. If you need to know which version a specific program is using, check that program's settings or documentation.

What if nvidia-smi shows a version but my software says CUDA is not found?

Your CUDA toolkit is installed but your software cannot find it. This usually means the software is looking in the wrong folder or your system path is not set up correctly. Reinstalling CUDA or the software itself usually fixes this. Check the software's documentation for CUDA setup instructions specific to your operating system.

Does updating my graphics driver change my CUDA version?

Updating the driver can change which CUDA version is available, but not always. Driver updates and CUDA updates are separate. You may need to update CUDA separately even after updating your driver. Check NVIDIA's website to see which CUDA versions your driver supports.

Why does my CUDA version say 12.0 but the software needs 12.1?

CUDA versions are not always backward compatible in the opposite direction. If software needs 12.1 and you have 12.0, you need to update CUDA. read the newer version from NVIDIA's website and reinstall. Uninstall the old version first to avoid conflicts.

Can I check CUDA version without a graphics card?

No. CUDA requires an NVIDIA graphics card to function. If you do not have one, CUDA cannot be installed or checked. Some cloud services and remote computers have NVIDIA cards, so you could check CUDA on those systems instead.