JetPack 6.2 / TensorRT 10.3 on Orin NX/Nano: engine runs but outputs are numerically wrong (flat planes) vs correct output on DGX (TRT 10.14

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
1

Hi NVIDIA team,

We are deploying a 3D reconstruction pipeline based on the VGGT model. The pipeline works as expected on our DGX Spark environment, but when we run the same PyTorch → ONNX → TensorRT workflow on Jetson Orin (NX/Nano), the inference results are incorrect. Specifically, instead of generating proper 3D surfaces, the model outputs flat planes; so the engine runs without crashing, but the output is silently wrong.

Suspected cause: We believe the issue stems from missing or incomplete operator and tactic support in JetPack 6.2’s TensorRT 10.3 on Orin. The features and kernels required by our model may not be fully implemented or optimized, leading to silent inference errors. On DGX, we have newer software (JetPack 7.0, TensorRT 10.14, CUDA 13.0) and off course more advanced hardware (Blackwell-class GPU) kconversion and inference work correctly. However, Jetson Orin (Ampere SM87) seems to lag behind in compatibility and numerical precision.

VGGT context: The VGGT model is a complex 3D neural network (~4.4 GB FP32) that requires robust operator support for multi-view depth and point cloud inference. JetPack 6.2/Orin appears to lack some of this essential support.

Experimental findings:

  • The model performs correctly on DGX, but fails on Orin.

  • DGX runs newer JetPack/TRT/CUDA versions, while Orin is limited to JetPack 6.2/TRT 10.3.

  • The incorrect “flat surface” output on Orin suggests missing tactic/operator support, not an export or build failure.

Question: Has anyone else encountered similar inference issues with VGGT-based or similar models on Jetson Orin, and are there recommended ways to improve operator/tactic support or workaround these limitations?

5

Hi,

Could you share a reproducible source with us so we can check further about the issue?
It will be good if you can attach a sample that generates the correct output (with onnxruntime?) and reproduces the issue (TensorRT 10.3).

Thanks.

6

Dear AstaLLL,

My colleague Kai has repeated to collect sample data with correct output and the reproduced issue. However he created a new post for this: TensorRT engine produces distorted outputs on Jetson Orin NX (JetPack 6.2 / Ampere) but matches ONNX on DGX Spark (JetPack 7.0 / Grace Blackwell).

We can close this topic and move towards the new post?

Walter

7

Hi,

Sure. We will follow up on this issue on the new topic directly.
Thanks.