`humble` image with cross architecture GPU support

[rtarun1]

Brief

• As GPU is being used in every part of robotics, its fair that we need a ROS docker image with GPU support out of the box without messing up any local deps, especially when it comes to Nvidia deps, which is often a headache.
• So we can have a humble image which builds on top of nvidia/cuda:12.2.2-devel-ubuntu22.04, instead of ubuntu:jammy, with no other change to our humble image.
• Reason to choose cuda-12.2: In my experience it seems to be in a sweet spot where old deps that rely on older versions of CUDA and also it covers any new work that we can implement.
• Pros:
  • Thanks to the amazing work by nvidia team, it comes out of the box with CUDA and CuDNN.
  • It works cross architecture, without any changes. Tested on ARM based Jetson Orin GPU and regular x86 based system with Nvidia 2060 mobile GPU
• Cons:
  • The images are 3x the size, for context, our image is 3.52GB, but the GPU image is 10.2GB

[soham2560]

shoukd a base image with nvidia/cuda:12.2.2-base-ubuntu22.04 and one with nvidia/cuda:12.2.2-devel-ubuntu22.04 be adopted then, will that be better?

[rtarun1]

shoukd a base image with nvidia/cuda:12.2.2-base-ubuntu22.04 and one with nvidia/cuda:12.2.2-devel-ubuntu22.04 be adopted then, will that be better?

Yup, that seems like a good idea, because building on top of nvidia/cuda:12.2.2-base-ubuntu22.04 only consumes 3.71GB

[soham2560]

also could you elaborate on the cross architecture part of the PR title

[rtarun1]

When building for ARM architecture, we typically base our images on docker.io/arm64v8/ubuntu:22.04, which forces us to maintain separate Dockerfiles for different architectures. Additionally, adding CUDA dependencies becomes even more complex, as it further branches away from our standard Humble build.

However, by using the official NVIDIA CUDA base images, such as nvidia/cuda:12.2.2-base-ubuntu22.04, the underlying architecture is handled automatically. This allows us to use a single unified Dockerfile for both x86 and ARM architectures, while seamlessly integrating CUDA dependencies.

[soham2560]

ooh really, the nvidia images handle it automatically? then same should be the case for ubuntu too right, hard to believe they aren't doing multi arch builds, for ubuntu that too, bit busy atm so not able to fully look into it, but let me know if you find the reason for this discrepancy