• getting started
• Jetson examples
• nvidia certification (no longer available)
• Jetson AI lab

• Getting Started with AI on Jetson Nano

• do not try to update with seeedstudio's firmware. It's not updateable after that
• instead use sdkmanager from nvidia:
  • do manual install. Automatic will try to update the seeedstudio stuff which will fail
  • manual install requires forced recovery mode which requires a USB connection
    • to do forced recovery with unit unplugged:
      • use a paper clip to press the recovery button
      • while pressing the recovery button apply power
  • attach USB cable to “device port.” For USB the cable used matters. I had luck using usbc ←→ usbc. Using usba ←→ usbc failed
  • 1/2 way through the install process, sdkmanager will want to reconfirm/switch how host is communicating to the target. At this point the target is booted into the new OS
  • at that point I switched to ethernet as both units were connected to my router. I'm not sure if sticking with usb might have worked or would have been faster

corrected from: https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html)

• install jetpack. (This is not the same as installing the OS for Jetson): sudo apt install nvidia-jetpack
• get version of CUDA: nvcc --version
• install cusparselt:

wget raw.githubusercontent.com/pytorch/pytorch/5c6af2b583709f6176898c017424dc9981023c28/.ci/docker/common/install_cusparselt.sh 
sudo -s # set CUDA version and install as root
export CUDA_VERSION=12.1 # as an example. Use the version returned from ''nvcc --version''
bash ./install_cusparselt.sh

• install correct numpy version: sudo pip install numpy==1.26.1
• download correct version of pytorch for jetpack version. Search for it at https://developer.download.nvidia.cn/compute/redist/jp/

sudo pip3 install --no-cache http://<URL TO PYTORCH WHL>

• onnxruntime for Jetson
• if onnxruntime fails with message Invalid argument. Specify the number of threads explicitly fix with one-time cli command: sudo nvpmodel -m 0 which forces the computer to with no throttling
• tegrastats might show load on GPU?
  • tegrastats | cut -f 14 -d " " shows % load on GPU