Please provide the following info (tick the boxes after creating this topic):
Software Version
DRIVE OS 7.0.3
other
Target Operating System
Linux
QNX
other
Hardware Platform
DRIVE AGX Thor Developer Kit (940-63960-0010-000)
DRIVE AGX Thor Developer Kit (940-63960-0012-000)
other
Host Machine Version
native Ubuntu Linux 24.04 Host installed with DRIVE OS Docker Containers
other
Issue Description
Not an issue. Just curious why our NUC has to have a GPU according to the requirements here (https://developer.nvidia.com/docs/drive/drive-os/7.0.3/public/drive-os-linux-installation/requirements.html).
Are we supposed to run inferences first on NUC docker to test and then copy it to Thor? What would be a standard workflow then?
Dear @quan.luo.101 ,
GPU is needed only if you want to run the application on host. If you plan to run on DRIVE, you can cross compile the application on docker and copy the binaries to target to run.
What about tensorrt engine? I think we’ll still need to build the tensorrt engine on Target Drive Thor?
Or we can build the engine on host inside the docker? (So that we might need GPU in this sense)
Dear @quan.luo.101 ,
TensorRT engine needs to be generated on target if you want to use it on target.
Thanks for replying in time!
So the host machine is just for flashing & crosscompiling CPU tasks? Are we supposed to develop on Thor directly since it seems that Drive Thor has only ~30GB mounted on root…
you can develop the application on docker and cross compile the application to generate binaries. The docker container has the necessary build environment to compile DRIVE OS applications (DW, TRT, CUDA, NvMedia etc).
If you want to perform inference, only the TRT engine needs to generated on Thor and use it in the application. We have trtexec tool in TRT to achieve this