Applies to: All models in Advantech MIC AI series

Problem:

When I check my system, I see that NVIDIA libraries like CUDA, cuDNN, and TensorRT are missing, or my AI application (using PyTorch, TensorFlow, YOLO) is not using the GPU and is running slowly. How can I correctly install the necessary NVIDIA Jetson SDK components?

Solution:

The NVIDIA SDK components (CUDA, cuDNN, TensorRT, etc.) are not included in the default BSP and must be installed separately. The recommended method is a two-stage process that involves flashing the Advantech BSP first, followed by installing the SDK components using the NVIDIA SDK Manager from a host computer.

Requirements:

• An Advantech Jetson device (the "target").
• A separate host computer running Ubuntu 22.04 (x86_64 architecture is preferred).
• A network connection where both the host PC and the target device are on the same subnet.

Step-by-Step Installation Guide:

Stage 1: Flash the Device with the Advantech BSP

1. Download the correct Advantech BSP that matches your specific device model and the desired NVIDIA JetPack version (e.g., JetPack 6.2).
2. Follow the SOP provided with the BSP to flash your device.
3. Ensure the device boots successfully after flashing is complete.

ℹ️ Note

Please check the BSP Support page for downloading your BSP.

Stage 2: Install NVIDIA SDK Components via SDK Manager

1. On your x86_64 host PC, download and install the NVIDIA SDK Manager.

   • You can download it from the official NVIDIA website: https://developer.nvidia.com/sdk-manager

2. Connect your target device to the same network as your host PC. You do not need to keep the micro-USB cable connected after the initial flashing is done.

3. Launch the SDK Manager on your host PC and log in.

4. In Step 01 (Hardware Configuration):

   • Uncheck the "Host Machine" box to reduce installation time.
   • From the "Target Hardware" dropdowns, select the correct product category (e.g., Jetson AGX Orin or Jetson Orin NX).
   • Select the "Target Operating System" and "JetPack version" to exactly match the Advantech BSP you flashed in Stage 1. This is critical for compatibility.
   • Enter the IP address, username, and password for your target device.

5. In Step 02 (Details and License):

   • IMPORTANT: Uncheck the box for "Jetson OS" (sometimes labeled "Jetson Linux"). You are only installing the "Jetson SDK Components".
   • Accept the license agreements and continue.

6. In Step 03 (Setup Summary):

   • The manager will begin downloading and then installing the selected SDK components onto your target device over the network.

7. Once the installation is complete, reboot your Jetson device. You can now verify that the CUDA, cuDNN, and TensorRT libraries are installed and available.

By following this procedure, you ensure that the operating system, kernel, and NVIDIA drivers are correctly aligned, preventing the most common installation and compatibility issues.