[Question] Has anyone successfully run GPUDrive on RTX 50 series (sm_120) GPUs?

[RTGTX7]

I’m currently trying to build and run GPUDrive on a new GPU from the RTX 50 series (specifically an RTX 5070 Ti). These GPUs belong to the Blackwell architecture family and use Compute Capability sm_120.

Here is what I observed:

• CPU backend: Everything builds and runs fine.
• CUDA backend: The project builds successfully with CUDA 12.4 (using stubs for libcuda.so), but when running, any CUDA kernel launch immediately crashes. There is no detailed error message other than the kernel failing.

---

Questions

1. CUDA toolkit version
   The documentation states GPUDrive supports CUDA 12.2 – 12.4.
   However, NVIDIA’s official docs say sm_120 (Blackwell) is only supported starting with CUDA 12.8.

   • Has GPUDrive ever been tested on sm_120 GPUs with CUDA 12.4?
   • Is the crash expected because CUDA 12.4 doesn’t fully support this architecture?

2. Real-world usage
   Has anyone (dev team or community) successfully run GPUDrive on RTX 50 series GPUs (e.g., 5070 Ti, 5080)? If yes, could you please share:

   • CUDA version
   • Driver version
   • Any special build flags (e.g., CMAKE_CUDA_ARCHITECTURES, PTX fallback)?

3. Recommended setup

   • If CUDA 12.4 cannot reliably support sm_120, should we move to 12.8+ despite the current documentation limiting support to <=12.4?
   • Is official CUDA 12.8 (or newer) support planned for GPUDrive?

---

My environment

• GPU: RTX 5070 Ti (sm_120, Blackwell)
• Driver: 570.xx (Open Kernel Driver)
• Base image: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
• OS: Ubuntu 22.04 (inside Docker)

Symptom: CPU runs correctly, but any CUDA kernel invocation crashes at runtime under this setup.

---

Summary

• CPU backend works fine.
• CUDA backend builds, but kernel launches crash immediately.
• Suspect this is due to missing sm_120 support in CUDA 12.4.

Looking for confirmation and recommendations:

• Has anyone run GPUDrive successfully on RTX 50 series yet?
• Should I switch to CUDA 12.8+ for these GPUs?

Thanks again for your time and for maintaining this project!