CUDA NVRTC User Guide (NVIDIA, v13.4 docs)

NVRTC is NVIDIA's runtime compilation library for CUDA C++. Primary use: it accepts CUDA C++ source as a character string and produces GPU executable code - PTX, cubin, or cuda_tile IR - loadable with cuModuleLoadData/cuModuleLoadDataEx, and linkable with other modules via nvJitLink or cuLinkAddData. The docs state the linking facility 'can often provide optimizations and performance not possible in a purely offline static compilation'. Supported platforms are Linux x86_64, Linux aarch64, Windows x86_64, and it 'does not depend on any other libraries or headers from the CUDA toolkit, and can be run on a system without a GPU'. The docs also flag a real operational hazard in a footnote: 'Changes in compiler optimizer heuristics in the newer NVRTC shared library may also potentially cause performance perturbations for generated code' - i.e. an NVRTC upgrade can silently move kernel performance with no source change. Used here for what NVRTC compiles TO and what its library boundary is; no speed claim.