Compatibility of CUDA-BEVFusion with NVIDIA Drive AGX

[Zizzi23]

Hi, First of all, thank you for your great work on the CUDA-BEVFusion project — it's impressive and exactly what I was looking for.

I noticed that in your repository, you mention that the model achieves 25 FPS on NVIDIA Orin using TensorRT with INT8 optimization. I'm currently working on a project using the NVIDIA Drive AGX platform (instead of Orin), and I'm considering adapting CUDA-BEVFusion for it.

I’d like to ask: Do you see any issues or limitations in running your model on NVIDIA Drive AGX (e.g. Xavier or Pegasus)? Would you recommend any adjustments for compatibility or performance optimization?

Thanks in advance for your help!

[hopef]

For the CUDA-BEVFusion project, currently only SM87 (Ampere, Orin), SM80 (Ampere, A100), and SM101 (Blackwell) are supported. Unfortunately, Xavier (Volta) does not support running this project as the primary limitation is the ampere MMA instructions in the libspconv library. Unless you replace the libspconv in CUDA-BEVFusion with a non-sparse convolutional model, such as pointpillars, this may require a lot of effort.