Hi NVIDIA team,
I am deploying a PillarNet-18 3D object detection model using libspconv (from Lidar_AI_Solution) on an aarch64 system with a GB10 (Blackwell, SM121) GPU.
The x86_64_cuda13.0/libspconv.so (v1.3.2) already includes SM120 SASS kernels (sm_80/86/90/110/120), confirmed via cuobjdump. However, the aarch64 prebuilt binaries do not include SM120:
| Binary | Architectures | SM120 |
|---|---|---|
x86_64_cuda13.0 (53 MB) |
sm_80, sm_86, sm_90, sm_110, sm_120 | Yes |
aarch64_cuda12.8 (27.8 MB) |
sm_101a only | No |
aarch64_cuda13.0 (24.3 MB) |
sm_110 only | No |
Both aarch64 binaries output the error:
“Blackwell kernels are not supported for this architecture. Please use CUDA 12.8 with SM100 or higher (except SM120).”
Running the demo on GB10 results in a segfault due to kernel incompatibility.
Environment
- GPU: NVIDIA GB10 (Blackwell, compute capability 12.1)
- Platform: aarch64 (ARM 64-bit)
- OS: Ubuntu 24.04.4 LTS
- CUDA: 12.8 (V12.8.93)
- Driver: 590.48.01 (supports CUDA 13.1)
- libspconv: v1.3.2
Request
Could you please provide an aarch64 prebuilt libspconv.so that includes SM120 kernels, matching the existing x86_64_cuda13.0 binary?
Related
- GitHub Issue: [Request] Add aarch64 libspconv binary with SM120 (Blackwell GB10) support · Issue #361 · NVIDIA-AI-IOT/Lidar_AI_Solution · GitHub
- GitHub Issue #341 (SM120 resolved for x86_64): Support SM 120 (blackwell) · Issue #341 · NVIDIA-AI-IOT/Lidar_AI_Solution · GitHub
- GitHub Issue #347 (SM120 resolved for x86_64): CUDA 13.0 SM_120 Support · Issue #347 · NVIDIA-AI-IOT/Lidar_AI_Solution · GitHub
Thank you!