Hi,
I am currently trying to build a calibration cache to use it subsequently on PX2. I follow this example:
https://devblogs.nvidia.com/int8-inference-autonomous-vehicles-tensorrt/
The question is whether I should use the version of TensorRT on my host machine as the version on PX2. Currently, they are different, namely, TRT 7.0.0.11-1+cuda10.2, and 4.0.0.8-1+cuda9.2. So do I need to downgrade TRT& Cuda on my host machine first!?
Thanks a lot in advance!
Hi @shd2020,
when developing on the host for later deploying on the target, you should always use the same version of software for best compatibility.
when installing the SDK using sdk manager it installs the correct (and same) versions on the host and on the target.
Please use the TensorRT and CUDA version that was installed by SDK Manager and not other manually installed version.
to reinstall the right version on your host please use SDK Manager and choose only HOST installation.
1 Like
Hi @shayNV,
Thank you very much for your prompt reply. I assume that you mean that I should install NVIDIA DRIVE™ OS 5.0.10.3 with NVIDIA DriveWorks SDK for NVIDIA DRIVE™ PX 2 on my host machine, right?
Yes
Hi @SivaRamaKrishnaNV,
Thank you very much. The problem is that my host machine is a server, so I’m afraid there might be a problem with enabling X11 on it. So I had been wondering whether I could just manually install TRT 4.0.0.8-1+cuda9.2 on my host instead so that it would match PX2. Thanks in advance for your advice on that!
Dear @shd2020,
Please install the setup on your host using sdkmanager. It installs the required CUDA and TRT versions. After getting the calibration cache on your host, you can copy it to DRIVE PX2 for reuse.