GPU for Blender Rendering

Hi, thanks for the great work on Kubric.

I built Kubric’s Blender Docker image using Blender 2.93 LTS with CUDA enabled. CUDA is detected correctly and the GPU allocates significant VRAM (~9–10 GB), but Cycles GPU rendering is extremely slow, significantly slower than CPU rendering.

This appears similar to issue [#224](https://github.com/google-research/kubric/issues/224). However, the solutions discussed there seem outdated and no longer work with modern GPUs (e.g., RTX 40-series). Hence I wanted to open a new issue.

I suspect this may be related to Blender 2.93 and the older CUDA toolchain being incompatible with newer GPUs, even though CUDA is detected.

Could you clarify:
- Whether GPU rendering with Blender 2.93 is still expected to work reliably in Kubric
- If upgrading to a newer Blender/CUDA combination is the recommended approach
- Or if there is a known supported configuration for GPU rendering

Any guidance would be greatly appreciated.
Thank you.

Start by reviewing the Kubric Blender Docker image configuration and the discussion in issue #224, focusing on the Blender 2.93 and CUDA versions mentioned here. Done should be a documented supported GPU-rendering configuration, or a confirmed explanation that this setup is not supported.

Written by the indexing model from the issue text.