OpenCV with CUDA

What you'll build

OpenCV and its contributed modules, built from source with CUDA, cuDNN and the CUDA backend for the DNN module, installed under /usr/local for C++ and Python. The steps work on other Ubuntu or Debian systems with NVIDIA GPUs, given a compatible CUDA, compiler and driver.

What the build includes

To add any of these, rebuild with extra CMake options; see OpenCV's configuration reference.

Before you start

• Ubuntu 24.04 as recorded; other releases need a compatible CUDA and compiler.
• An NVIDIA GPU with CUDA support, and its driver installed.
• About 5 GB of free disk space for the build.
• Root or sudo access.

1. Install the dependencies

1. Install the CUDA Toolkit and cuDNN from NVIDIA, and CMake, GCC 12 and Python 3 from Ubuntu:

   Package	Version	Where to get it
   CUDA Toolkit	A CUDA 12 release compatible with your OS, compiler and GPU	NVIDIA CUDA downloads
   cuDNN	8.9.2 or compatible	NVIDIA cuDNN (free developer account)
   CMake	3.15 or later	sudo apt install cmake
   GCC and G++	12 recommended	sudo apt install gcc-12 g++-12
   Python 3	With headers and NumPy	sudo apt install python3 python3-dev python3-numpy

2. Install the image and video libraries, which add support for common formats:

   sudo apt install build-essential cmake git pkg-config \
       libgtk-3-dev libavcodec-dev libavformat-dev libswscale-dev \
       libv4l-dev libxvidcore-dev libx264-dev libjpeg-dev libpng-dev \
       libtiff-dev gfortran openexr libatlas-base-dev libtbb-dev \
       libdc1394-dev libopenexr-dev libgstreamer-plugins-base1.0-dev \
       libgstreamer1.0-dev libwebp-dev

2. Find your GPU's compute capability

The build needs your GPU's compute capability.

1. Check your GPU model with nvidia-smi.
2. Look up its compute capability on NVIDIA's CUDA GPUs page.
3. Note the value, for example 9.0 for H100 and H200, 8.6 for RTX 3090, or 7.5 for RTX 2080.

3. Get the source

1. Make a working folder:

   mkdir opencv
   cd opencv

2. Download OpenCV and its contributed modules:

   curl -Lo opencv-4.12.0.tar.gz https://github.com/opencv/opencv/archive/4.12.0.tar.gz
   curl -Lo opencv_contrib-4.12.0.tar.gz https://github.com/opencv/opencv_contrib/archive/4.12.0.tar.gz

3. Extract both:

   tar xf opencv-4.12.0.tar.gz
   tar xf opencv_contrib-4.12.0.tar.gz

4. Make the build folder:

   mkdir build
   cd build

4. Configure with CMake

cmake -DOPENCV_EXTRA_MODULES_PATH=../opencv_contrib-4.12.0/modules \
    ../opencv-4.12.0 \
    -D WITH_CUDA=ON \
    -D WITH_CUDNN=ON \
    -D OPENCV_DNN_CUDA=ON \
    -D CMAKE_C_COMPILER=gcc-12 \
    -D CMAKE_CXX_COMPILER=g++-12 \
    -D CUDA_ARCH_BIN=9.0

The build installs under /usr/local in step 6, where the system package manager doesn't track it, as OpenCV notes. To keep it separate, add -D CMAKE_INSTALL_PREFIX= with another location to this command.

The configuration summary shows CUDA and cuDNN as found. If either says NO, see Troubleshooting before compiling.

5. Compile

This uses 60 parallel jobs; set -j to your core count, or use -j $(nproc) to use them all. It takes 15–60 minutes.

cmake --build . -j 60

6. Install

1. Install:

   sudo cmake --install .

   Programs go in /usr/local/bin, libraries in /usr/local/lib and headers in /usr/local/include/opencv4. The Python module goes in the path CMake reported; check which copy Python loads with python3 -c "import cv2; print(cv2.__file__)".

2. Refresh the library cache so the system finds the new libraries:

   sudo ldconfig

7. Verify

Check the build from Python, then from C++.

Python

python3 -c "import cv2; print(f'OpenCV version: {cv2.__version__}')"

OpenCV version: 4.12.0

CUDA support

python3 -c "import cv2; print(f'CUDA devices: {cv2.cuda.getCudaEnabledDeviceCount()}')"

The count is greater than 0.

Build information

python3 -c "import cv2; print(cv2.getBuildInformation())" | grep -A 5 "NVIDIA CUDA"

It shows your CUDA version and GPU architecture.

C++

1. Create test_opencv.cpp:

   #include <opencv2/opencv.hpp>
   #include <opencv2/cudaarithm.hpp>
   #include <iostream>
   
   int main() {
       std::cout << "OpenCV version: " << CV_VERSION << std::endl;
       std::cout << "CUDA devices: " << cv::cuda::getCudaEnabledDeviceCount() << std::endl;
       return 0;
   }

2. Compile and run it:

   g++ test_opencv.cpp -o test_opencv $(pkg-config --cflags --libs opencv4)
   ./test_opencv

It prints version 4.12.0 and a CUDA device count above 0.

Troubleshooting

CMake reports CUDA: NO

• Check CUDA is installed: nvcc --version.
• Make sure /usr/local/cuda/bin is on your PATH.
• Check the CUDA libraries are in /usr/local/cuda/lib64.

CMake reports cuDNN: NO

• Check cuDNN is in /usr/local/cuda/include and /usr/local/cuda/lib64.
• Make sure the cuDNN version matches your CUDA toolkit.
• Set CUDNN_INCLUDE_DIR and CUDNN_LIBRARY explicitly in CMake.

Errors about the CUDA driver version

• nvidia-smi shows both the driver version and the CUDA version it supports.
• Make sure the driver supports your CUDA toolkit.
• Update it with sudo apt install nvidia-driver-xxx, replacing xxx with the right version.

Python can't load libopencv_core.so.4.12

• Run sudo ldconfig.
• Check that /usr/local/lib is listed in /etc/ld.so.conf.d/.
• Check Python's path: python3 -c "import sys; print(sys.path)".

The build fails with compiler errors

• Check every prerequisite is installed.
• Try another GCC version (9, 10, 11 or 12).
• Check there's about 5 GB of free disk space.
• Empty the build folder and configure again: rm -rf * && cmake ..., from inside opencv/build only.

Testing and references

The original notes record Ubuntu 24.04, CUDA 12.0, cuDNN 8.9.2 and an H200. That exact environment hasn't been reconfirmed, and CUDA 12.0's supported distributions don't include Ubuntu 24.04. Check the toolkit, compiler, driver and build logs before treating it as a supported recipe. Selected commands were checked against documentation in September 2026; the build wasn't rerun.

• OpenCV documentation
• OpenCV CUDA module
• OpenCV on GitHub