How to get CUDNN and CUDA working on Jetpack 5.1.2

I have reflashed the Orin using the SDK manager and I also installed cuda 11.4 and jetpack 5.1.2 trough the same process. When I try nvcc --version it returns that it has cuda 11.4.

However, when I tried to run the cuda samples it told me that I got this error: VulkanBaseApp.cpp:30:10: fatal error: GLFW/glfw3.h: No such file or directory. So I did “sudo apt install sudo apt-get install libglfw3 libglfw3-dev”. Afterwards the cuda samples did work. So now CUDA seems to be working.

Now I wanted to test if CUDNN was properly installed. So I tried to run this command “cat /usr/include/cudnn.h | grep CUDNN_MAJOR -A 2”, which should detect the CUDNN version, but it returns nothing, like there is a new terminal line immediately under it.

When I try to run samples/run yolov8 models (from the ultralytics library) I get a worse/slower performance than on a laptop that is using a non powerful CPU. This lead me to believe that CUDNN is not properly installed.

Hi,

The version info is included in cudnn_version.h rather than cudnn.h.

$ cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2
#define CUDNN_MAJOR 8
#define CUDNN_MINOR 6
#define CUDNN_PATCHLEVEL 0
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

/* cannot use constexpr here since this is a C-only file */

To deploy a DNN model, please ensure your frameworks are built with CUDA support.
Usually, the package directly from the pip only has CPU support.

Below are some common DNN frameworks of Jetson for your reference:

Thanks.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.