Predictions on Jetson Xavier NX wrong while working on NVIDIA T1000 laptop gpu

Description

We have a program that converts a onnx to a tensorrt engine and runs a prediction. This program works on a T1000 laptop gpu running in a docker container. But returns wrong output results on a Jetson Xavier NX. Besides the platform architecture, everything is more or less the same.

I also noticed that the engine on the Jetson Xavier NX is only half the size of that running on the laptop.

More info and steps to reproduce can be found down below.

Many thanks in advance

Environment

Host machine:

TensorRT Version: 8.5.2
GPU Type: NVIDIA QUADRO T1000
Nvidia Driver Version: 517
CUDA Version: 11.7
CUDNN Version: 11.8
Operating System + Version: Windows 11
Python Version (if applicable): /
TensorFlow Version (if applicable): /
PyTorch Version (if applicable): /
Baremetal or Container (if container which image + tag): Container (see dockerfile in github)

Jetson:

TensorRT Version: 8.4.1.5
GPU Type: Jetson Xavier NX
Nvidia Driver Version: ??
CUDA Version: 11.4
CUDNN Version: 8.4.1
Operating System + Version: Jetpack 5.2
Python Version (if applicable): /
TensorFlow Version (if applicable): /
PyTorch Version (if applicable): /
Baremetal or Container (if container which image + tag): baremetal

Relevant Files and Steps To Reproduce

See GitHub: https://github.com/SirMomster/tensorrt-recognition-jetson-reproduce

Hi,

We are moving this post to the Jetson Xavier NX forum to get better help.

Thank you.

Hey spolisetty,

We did some further testing and found out that the Jetson doesn’t (directly) seem to be the problem.

After these further tests, we found out that the models that we are using which are Scene Text Detection models work fine on Tensorrt 8.5.2.2 with CUDA 11.8 and cudNN 8.7. They do not work on Tensorrt 8.5.1.7 nor on 8.4.1.5 (which is currently included in Jetpack 5.0.2).

Which is great, but as far as we know, there is no current way to update the Tensorrt version on the Jetson Xavier NX. As these are prepackaged and managed by the SDK manager for Jetpack.

Is there anything on the roadmap to support Tensorrt 8.5.2.2 in the near future on the Jetson’s?

(Side note: I removed the mentioned GitHub repository as the problem wasn’t related to it)

Kind regards,
Mitch

Hi,

TensorRT 8.5.2 is included in our upcoming release (JetPack 5.1) which will be available soon.
You can find the roadmap below:

Thanks.

Hey,

Many thanks for the reply and the additional information regarding the roadmap.

Kind regards,
Mitch

Hi,

JetPack 5.1 is released.
Please check if TensorRT 5.1 can fix your issue.

Thanks.