CUDA 13 Support for libc++ on x86_64

Accelerated Computing CUDA CUDA Setup and Installation
1

Hi team Nvidia,

I’m currently working on a project that is moving from CUDA 11.8 to CUDA 13 to target Blackwell (sm_100/120). My toolchain is entirely LLVM/Clang based, but I’ve hit a hard blocker in the CUDA headers.

Specifically, crt/host_defines.h is throwing: error: "libc++ is not supported on x86 system"

The Situation:

  • Host: x86_64 Linux.

  • Target: Blackwell.

  • Constraints: I’m tied to libc++ for ABI compatibility with several other non-CUDA dependencies. Moving back to libstdc++ isn’t feasible without rebuilding the entire world, which isn’t an option here.

My Questions:

  1. Are there any official flags or “preview” modes in CUDA 13 to allow libc++ on x86_64?

  2. If I define _ALLOW_UNSUPPORTED_LIBCPP to bypass this, are there known breaking issues?

  3. what is the recommended workaround for teams tied to the LLVM ecosystem who need to target Blackwell?

Thanks