ZED SDK 5.4.0 on DGX Spark (GB10) fails NEURAL depth model optimization and enters infinite optimization loop

Software
1

Hi Stereolabs team,

I’m testing the new DGX Spark beta support introduced in ZED SDK 5.4.0 and have encountered an issue with NEURAL depth model optimization.

Hardware

  • NVIDIA DGX Spark (GB10)

  • System reports: Dell Pro Max with GB10 FCM1253

  • Driver Version: 580.142

Software

  • Ubuntu 24.04.4 LTS

  • ZED SDK 5.4.0

  • Fresh installation

Issue

After installing the SDK and attempting to run any application that uses NEURAL depth mode, the SDK prompts for AI model optimization.

The optimization process starts but eventually fails with the message:

Failed to optimize model. NEURAL mode won’t be available.

After this failure, the application immediately triggers the same optimization prompt. The model never successfully optimizes, and the system appears stuck in a loop where optimization is repeatedly requested but always fails.

Steps to Reproduce

  1. Install ZED SDK 5.4.0 on DGX Spark (GB10).

  2. Launch an application that uses NEURAL depth mode (or attempt model optimization manually).

  3. Optimization begins.

  4. Optimization fails with:

    • “Failed to optimize model. NEURAL mode won’t be available.”

  5. SDK requests optimization again and repeats the same failure.

I have not yet found a way to complete model optimization successfully. Happy to provide diagnostic logs if needed.

Thanks.

2

Hi @ShashankVSS
DGX Spark support is still experimental and not completely operational.
Your report will help us improve the compatibility of the ZED SDK with this platform.

Can you please run the ZED Diagnostic tool and share the JSON report that it generates?

3

@ShashankVSS In addition to the ZED DIagnostic tool report, pls also post the output of these commands to show TensorRT library version

  • ls /usr/local/cuda/lib64/ | grep libnvinfer
  • ldd /usr/local/zed/lib/libsl_ai.so | grep libnvinfer

While being experimental, this is definetly not expected and can’t be reproduced on our side currently.

4

Hi Myzhar, hi adujardin,

Thanks for the follow-up.

I ran the ZED Diagnostic tool and attached the generated JSON report.

Here are the requested TensorRT-related command outputs:

ls /usr/local/cuda/lib64/ | grep libnvinfer
libnvinfer_builder_resource_ptx.so.10.16.1
libnvinfer_builder_resource_sm100.so.10.16.1
libnvinfer_builder_resource_sm110.so.10.16.1
libnvinfer_builder_resource_sm120.so.10.16.1
libnvinfer_builder_resource_sm75.so.10.16.1
libnvinfer_builder_resource_sm80.so.10.16.1
libnvinfer_builder_resource_sm86.so.10.16.1
libnvinfer_builder_resource_sm89.so.10.16.1
libnvinfer_builder_resource_sm90.so.10.16.1
libnvinfer_plugin.so
libnvinfer_plugin.so.10
libnvinfer_plugin.so.10.16.1
libnvinfer.so
libnvinfer.so.10
libnvinfer.so.10.16.1

ldd /usr/local/zed/lib/libsl_ai.so | grep libnvinfer
        libnvinfer.so.10 => /usr/local/cuda/targets/sbsa-linux/lib/libnvinfer.so.10 (0x0000f8ac43530000)
        libnvinfer_plugin.so.10 => /usr/local/cuda/targets/sbsa-linux/lib/libnvinfer_plugin.so.10 (0x0000f8ac402b0000)

Please let me know if there are any other logs or environment details that would help.

ZED_Diagnostic_Results.json (6.4 KB)

5

Thanks, nothing stands out. The only diff I see is driver version but I doubt this has an impact (The test board I have is 580.159, you have 580.142), you can try updating just in case.

You can also try to run the optimized NEURAL *LIGHT model we have run to test

mkdir -p /usr/local/zed/resources ; wget https://download.stereolabs.com/ai/engine/.neural_depth_light_5.3.model_optimized-fbcbl-1-121-13000-101601-125-256-24576-48-100-512-48-1-0e55-512 -P /usr/local/zed/resources/

Then run ZED Depth Viewer or a sample

6

I’ll try to update my driver to see if that’ll help with this. I tried the command you gave but it still has issues with the model optimization failing. I’ll let you know if the driver update fixes the issue, otherwise, we’ll wait for another future release.

Thanks for your help!

7

Sorry I just realized it was NEURAL LIGHT I sent, don’t hesitate to give it a try !

8

Hey adujardin and myzhar,

Sorry for the late reply. I was able to update the driver to 580.159 and also tested the NEURAL LIGHT model you provided.

It looks like ZED Depth Viewer and ZED Studio still attempt to optimize the standard NEURAL model on startup, which either crashes during optimization or gets stuck repeatedly showing the message:

“NEURAL depth mode requires model optimization. This is done only once and will take a few minutes to complete. Do you want to proceed?”

However, if I manually select NEURAL LIGHT in ZEDfu, it runs correctly without any issues.

Would it be possible to get downloadable optimized versions of the NEURAL and NEURAL PLUS models as well? I’d like to test whether the problem is specific to the optimization step rather than the models themselves.

Thanks again for all your help!

Shashank