Downgrading Jetson Orin Nano from JetPack 6.0 to 5.1.3

Issue Overview

Users are experiencing difficulties downgrading their Nvidia Jetson Orin Nano development board from JetPack 6.0 to 5.1.3. The primary motivation for this downgrade is to resolve compatibility issues with PyTorch, which is not supported in CUDA 12.2 (included in JetPack 6.0). Users have reported problems with the SDK Manager not showing older versions, installation failures, and difficulties in flashing the QSPI to an older version. Additionally, even after installing CUDA 12.4 on JetPack 6.0, nvidia-smi still shows 12.2, and PyTorch for CUDA 12.4 is not functioning correctly.

Possible Causes

1. SDK Manager Limitations: The latest SDK Manager versions may not support older JetPack versions on newer Ubuntu host systems.

2. Firmware Incompatibility: The current firmware (version 36.3) may not be compatible with older JetPack versions.

3. Host OS Compatibility: Newer Ubuntu versions (e.g., 22.04) may not support the installation of older JetPack versions.

4. Hardware Detection Issues: The host system may fail to detect the Jetson board during the flashing process.

5. CUDA Version Conflicts: Inconsistencies between installed CUDA versions and those reported by the system.

6. PyTorch Compatibility: PyTorch versions may not be properly recognizing the installed CUDA version.

Troubleshooting Steps, Solutions & Fixes

1. Use an Older Ubuntu Host:

   • Install Ubuntu 18.04 on the host PC.
   • Download and install SDK Manager 2.0 or 2.1.

2. Try Archived Versions:

   sdkmanager --archived-versions
   

   Note: This may not work on Ubuntu 22.04.

3. Manual Installation:

   • Download the following packages from the NVIDIA Developer website:
     • Jetson Linux Driver Package (BSP)
     • Sample Root Filesystem
   • Extract the system image and use the initrd command to flash the device:

     sudo ./flash.sh p3768-0000-p3767-0005-a0-qspi internal
     

   Refer to the NVIDIA Jetson Linux Developer Guide for detailed steps.

4. QSPI Flashing:

   • Download firmware version 35.3.1.
   • Extract the file and navigate to the folder.
   • Put the Jetson in recovery mode.
   • Run the following command:

     sudo ./flash.sh p3768-0000-p3767-0005-a0-qspi internal
     

   Note: If the board is not detected, verify the USB connection and recovery mode status.

5. CUDA Version Management:

   • If CUDA 12.4 is installed but nvidia-smi shows 12.2, check the system PATH and LD_LIBRARY_PATH.
   • Ensure that the correct CUDA version is being used by PyTorch:

     import torch
     print(torch.version.cuda)
     

6. PyTorch Installation:

   • Install PyTorch versions compatible with the installed CUDA version.
   • If PyTorch is not showing as +cu, verify that it was installed with CUDA support.

7. Alternative Approach:

   • Consider installing only the Linux image without CUDA drivers from the SDK Manager.
   • Manually install the required CUDA version afterward.

8. Troubleshooting Installation Failures:

   • If installation fails at 96% or 98%, try clearing the SD card and reformatting it.
   • Ensure stable power and USB connections during the flashing process.
   • Check for any error logs or messages during the installation process.

9. Consult Official Documentation:

   • Refer to the NVIDIA Jetson Linux Developer Guide for the most up-to-date and comprehensive instructions on flashing and system recovery.

If these steps do not resolve the issue, consider reaching out to NVIDIA support or posting in the official Jetson developer forums for further assistance.