Cannot upgrade AGX Xavier : Stuck on Jetpack 4.6 - Options for upgrading libraries without OEM BSP support?

Robotics & Edge Computing Jetson Systems Jetson AGX Xavier
, , , , , ,
1

Hi everyone,

I’m working with a Jetson AGX Xavier (tegra194) device with the following software stack:

Device Information:

  • JetPack: 4.6 (L4T 32.6.1)
  • CUDA: 10.2.300
  • cuDNN: 8.2.1.32
  • TensorRT: 8.0.1.6
  • VPI: 1.1.15
  • OpenCV: 4.1.1 (without CUDA)
  • OS: Ubuntu 18.04 (aarch64)

Issue: This device is tied to an OEM-provided BSP , and the OEM(SmartCow AI Technologies) has stopped releasing updates for newer JetPack versions.

This has become a bottleneck since I need newer TensorRT, VPI, cuDNN, and CUDA versions. Notably, newer TensorRT and VPI releases have much better support for DLA, PVA etc., but on JetPack 4.6 I’m unable to fully utilize these hardware engines despite them being present on the device.

What I’m trying to understand:

  1. What is the highest version of JetPack supported for this device?
  2. (i) Is it feasible to migrate this device to a newer JetPack (e.g., 5.x) without OEM BSP support? (ii). If not, are there recommended approaches to selectively upgrade the compute stack (CUDA, TensorRT, etc.) while staying on JetPack 4.6?
  3. Has anyone successfully used:
  • NVIDIA APT repos to upgrade components beyond what ships with JP4.6?
  • Containers (NGC / L4T-based) to work around BSP limitations?
  • Custom BSP adaptation or partial flashing strategies?
  1. If upgrading the full stack isn’t possible, how can i use the newer versions of TensorRT, VPI; basically use the HW engines better?

Constraints:

  • I may not have full control over low-level BSP components, i.e., bootloader, kernel etc.
  • Stability is important

I need to ascertain what would be the least disruptive path to get access to newer inference/runtime stacks on this hardware.

Any guidance, especially from people who have dealt with OEM-locked Jetson platforms, would be very helpful.

Thanks in advance.
Regards.

2

We’re not able to help due to that required the custom BSP support to do the upgrade.
Those newer TensorRT. VPI…etc are binded with the newer kernel/low layer drivers., thus can’t be upgraded individually.

3

Thank you for clarifying the limitation.

I understand that the newer TensorRT, VPI, CUDA stack is tied to the underlying BSP, kernel, and low-level drivers, so a direct in-place upgrade is not possible on this platform.

That said, are there any other practical options worth considering on my side before concluding that this is not feasible? For example, would running a newer software stack inside a Docker container on the Jetson device help in any way, even if the host BSP remains on JetPack 4.6?

This hardware is a significant investment for me, and I would like to avoid letting it go to waste if there is any realistic path forward. I have already extended it as far as possible with the current stack, so I am trying to list every viable option before giving up on it.

If there are any supported workarounds, partial upgrade paths, or recommendations for getting better use out of the existing DLA, PVA, TensorRT capabilities on JetPack 4.6, I would appreciate your guidance.

Regards.

5

Hi,
The latest Jetpack release for Xavier is 5.1.6 r35.6.4. If you use AGX Xavier developer kit, you can follow the steps to re-flash the system:
Quick Start — NVIDIA Jetson Linux Developer Guide 1 documentation

Can download the packages from

Jetson Linux | NVIDIA Developer

If you use a custom board, would need to consult with the board vendor to get support.