"Installing TensorFlow with GPU support on Jetson Orin Nano (JetPack 6.0.1 / R36.4.7)"

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
,
1

I’m trying to get TensorFlow running with GPU acceleration on a Jetson Orin Nano Super, using JetPack 6.0.1 (Release R36.4.7). I’ve manually verified and linked all required CUDA libraries (libcublas, libcufft, libcurand, etc.) under /usr/local/cuda-12.6/lib64, and set LD_LIBRARY_PATH accordingly.

Despite this, TensorFlow still outputs: Cannot dlopen some GPU libraries. Skipping registering GPU devices...

Has anyone successfully installed a compatible TensorFlow wheel or container for JetPack 6.0.1 on ARM64? Any working builds or setup scripts would be greatly appreciated.

2

*** Please note that this reply is generated by LLM automatically ***

To get TensorFlow running with GPU acceleration on a Jetson Orin Nano Super using JetPack 6.0.1 Release R36.4.7, you’ll need to ensure that you have the correct CUDA libraries installed and linked. Here’s a step-by-step guide to help you install a compatible TensorFlow wheel or container for JetPack 6.0.1 on ARM64:

Prerequisites:

  1. Ensure that you have the latest version of JetPack 6.0.1 installed on your Jetson Orin Nano Super.
  2. Verify that you have the necessary CUDA libraries installed, including libcublas, libcufft, and libcurand.
  3. Make sure that the CUDA libraries are properly linked under /usr/local/cuda/lib64.

Installing TensorFlow:

  1. Option 1: Install TensorFlow using pip

    • Install the TensorFlow wheel using pip: pip3 install tensorflow
    • However, this method may not provide the optimal performance, as the pre-built wheel may not be optimized for the Jetson Orin Nano’s GPU architecture.

  2. Option 2: Build TensorFlow from source

    • Clone the TensorFlow repository: git clone https://github.com/tensorflow/tensorflow.git
    • Navigate to the TensorFlow directory: cd tensorflow
    • Configure the build: ./configure
    • Build TensorFlow: bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
    • Install the built TensorFlow package: pip3 install /tmp/tensorflow_pkg/tensorflow-*.whl

  3. Option 3: Use a pre-built TensorFlow container

    • Pull the TensorFlow container from NVIDIA’s NGC registry: docker pull nvcr.io/nvidia/tensorflow:20.03-tf1-py3
    • Run the container: docker run --runtime=nvidia --rm -it nvcr.io/nvidia/tensorflow:20.03-tf1-py3

Comparison of Options:

  • Option 1 (pip install): This method is the easiest, but it may not provide the best performance, as the pre-built wheel may not be optimized for the Jetson Orin Nano’s GPU architecture.
  • Option 2 (build from source): This method provides the best performance, as you can optimize the build for the Jetson Orin Nano’s GPU architecture. However, it requires more effort and expertise.
  • Option 3 (pre-built container): This method provides a good balance between ease of use and performance. The pre-built container is optimized for the Jetson Orin Nano’s GPU architecture, and it’s easy to use.

Working Builds or Setup Scripts:

  • NVIDIA provides pre-built TensorFlow containers for JetPack 6.0.1 on ARM64, which can be found on their NGC registry.
  • You can also use the TensorFlow build scripts provided by NVIDIA to build TensorFlow from source.

Conclusion:

To get TensorFlow running with GPU acceleration on a Jetson Orin Nano Super using JetPack 6.0.1 Release R36.4.7, you can use one of the three options mentioned above. Option 2 (build from source) provides the best performance, but it requires more effort and expertise. Option 3 (pre-built container) provides a good balance between ease of use and performance.

*** Please give the thumbs up if you get answers you like. Or provide feedback to help us improve the answer. ***

5

Hi,

Do you use JetPack 6.2.1+r36.4.7 or JetPack 6.0.1?

We don’t have a prebuilt TensorFlow package for the JetPack 6.2 environment.
You will need to build it from the source.

Thanks.

6

Is there a reliable guide somewhere? I’m using JetPack 6.2.1, but so far I haven’t found any instructions that really help. Even a Docker container could not be installed because it requires driver version 560.x, but only 540.4.0 is installed on the Jetson.

7

Hi,

We have a package that was built for JetPack 6.1:

https://developer.download.nvidia.com/compute/redist/jp/v61/tensorflow/

Would you might to give it a try?

Thanks.