There's an extended compute capability 9.0a for Hopper that was introduced in Cuda 12.0: https://docs.nvidia.com/cuda/archive/12.0.0/cuda-compiler-driver-nvcc/index.html#gpu-feature-list E.g. Cutlass leverages it: 5f13dcad78/python/cutlass/emit/pytorch.py (L684) This adds it to the list of permitted architectures to use in `cpp_extension` directly. Pull Request resolved: https://github.com/pytorch/pytorch/pull/110587 Approved by: https://github.com/ezyang
5f13dcad78/python/cutlass/emit/pytorch.py (L684)