If we set CMAKE_CUDA_ARCHITECTURES OFF, this will generate PTX only for the lowest supported compute capability, and SASS will be generated by the driver JIT for the actual GPU on the user's system at runtime. This may not be what you want if you're distributing precompiled binaries and want them to be runnable out of the box on all supported GPUs.
If we set
CMAKE_CUDA_ARCHITECTURES OFF, this will generate PTX only for the lowest supported compute capability, and SASS will be generated by the driver JIT for the actual GPU on the user's system at runtime. This may not be what you want if you're distributing precompiled binaries and want them to be runnable out of the box on all supported GPUs.