I’m working withe Jetson Nanos and Xavier NX and I was happy to see the progress of JetPack components in the JP5 Developer preview. There are some practical things like Ubuntu 20.04 support or the important update of VPI to v2.0. But the release notes at the VPI 2.0 web page makes me wonder: VPI 2.0 dropped support for Jetson Nano, TX2 and TX2 NX.
Maybe I have not yet enough experiences with Jetson device developments. But from my point of view I see following:
- VisionWorks is deprecated and the web page said we have to use VPI in future.
- VPI 1.2 is incomplete in the point of interoperability with Python (we see the good progress of this point in VPI 2.0).
- VPI 1.2 doesn’t have any mathematical image algorithms like VisionWorks (using C/CUDA we have to program this by yourself, native Python is to slow in most cases.)
- In JP5 we can now use Numba or PyTorch together with all backends by using the same memory place of image data without memcopy. In JP 4.x Numba and PyTorch-Cuda was not implemented to replace the missing mathematical image algorithms in VPI 1.2
For me and my colleagues is Python as development platform important to get a fast time to market / proof of concept to market. We develop special systems for research topics with one or a hand full units only. No mass products.
It seems, that NVIDIA starts with JP5 to deprecate Jetson Nano and TX2 (NX) support. Maybe since the Maxwell and Pascal architecture and so the possible CUDA versions.
Could we get an answer, what is planned for the future? Should we stop using Nano / TX2 developments? Which devices will be good supported in 2023/24? Or are there replaces of Nano devices in the same price region planned?