DRIVE OS Version: 7.0.3

In the DriveOS release documentation, there’s mention in Chapter 12 and elsewhere to install Nvidia Driver 570, specifically from https:∕∕us.download.nvidia.com∕tesla∕570.124.06∕NVIDIA-Linux-x86_64-570.124.06.run.

Is this just a recommendation since it includes Cuda 12.8? I assume this would have a specific minor version of cuda 12.8 along with other tools.

We want to know if this is a hard requirement for development environment from a safety standpoint or if it was just a recommendation for convenience. Our understanding is that for vehicle safety development, the development environment must also be well documented and controlled.

One example situation that will come up is that we may have newer GPUs like Pro Blackwell series which we’ve found some issues running on versions below 580 along with lack of support for MIG until that version. Even Thor runs driver version 581.0 with the Thor specific Cuda 12.8.

Is there an official recommendation on this topic? What if we developed on driver 590.x, but used Cuda library 12.8.x? Is there a specific Cuda minor version to use or latest 12.8.x is okay too?

Dear @ewu2 ,
The DOS 7.0.3 docker container has CUDA 12.8 already. You can install 590.x driver and run CUDA samples on docker.

Thanks @SivaRamaKrishnaNV

Outside of that container, can we use any recent CUDA 12.8 with any graphics driver that supports it as part of development process on a host system or as part of our CI to build and release artifacts to run on Thor?

I just want to make sure if we roll our own docker container or development machines manually install CUDA 12.8 without a container but use a newer driver version, that it’s a supported workflow for developing for Drive Thor and there are no safety concerns.

Dear @ewu2 ,
The CUDA 12.x you install on x86 is not same as the one we bundled with DRIVE OS . We suggest you to use the CUDA that is bundled with DRIVE OS docker and other packages like DW and TensoRT are tied to it.

Thank you for the clarification.