I am not sure I understand something: can I take any python program that runs on Raspberry Pi and run it as is on the Jetson Nano? Will it utilize the Jetson capabilities?
I assume the Jetson Nano has many GPUs, so they need to be utilize somehow.
Is there a driver or something that need to be done before taking any Python (like CV2) and use it on Jetson?
Just some random info on the topic:
- Jetsons have a single GPU, but each GPU has many CUDA cores.
- Any program which can run on an RPi probably (perhaps after a few tweaks) could run on a Jetson. However, that doesn’t cause the GPU to automatically be used.
- In no case would a non-GPU program from an RPi use the GPU on a Jetson without the programming being ported to use GPU.
- CUDA and GPU-based software must be developed separately. If an RPi uses CPU and not GPU, then the same program would also not use GPU on a Jetson unless it is ported/modified.
- Any program which does use CUDA (or GPU cores) would need to work with that particular release of CUDA on the Jetson. For example, if a program uses CUDA 10, and both PC and Jetson support CUDA 10, then the program can be made to work on both. Possibly one would need to recompile the software for the CUDA arch, but that’s trivial. A developer has the ability to compile for multiple CUDA architectures at the same time if desired. So far CUDA 11 has not been released for Jetsons. Older Jetson models could possibly never get CUDA 11 due to some products being end of life.
- PCs tend to have available a wider range of CUDA releases. Jetsons tend to support only a version 10.x CUDA. See the release notes for the flash software to know which CUDA release would be installed to the Jetson. Something using CUDA 11 on a PC would not work with a Jetson running CUDA 10.
- Software such as OpenCV probably requires a particular CUDA release to work.
- Some AI software is designed to use
nvidia-smi
to find eligible GPUs. This is a query tool which works over PCI/PCIe. Jetsons directly integrate a GPU to the memory controller and do not use PCI/PCIe, and thus any such software would require a workaround to tell it how to find the GPU.
2 Likes
This topic was automatically closed 2 days after the last reply. New replies are no longer allowed.