cccl

cccl is NVIDIA's official bundle of the three core C++ libraries for CUDA: Thrust, CUB, and libcudacxx. In plain terms, if you write code that runs on an NVIDIA GPU, these are the building blocks (sort, reduce, scan, plus a standard library that works on the device) so you do not hand-roll GPU kernels from scratch. It is Apache-2.0 and ships with the CUDA Toolkit.

Thrust is the high-level layer: parallel algorithms with an interface close to the C++ standard library, portable across GPU and multicore CPU backends. CUB is the low-level, speed-of-light layer for people writing custom kernels. libcudacxx is the CUDA C++ standard library. Most CUDA developers already use these without thinking about it.

There is nothing to decide here. If you write CUDA C++, you use CCCL. It is maintained by NVIDIA, unified into one repo, and free. The only catch is the obvious one: it is CUDA, so it is NVIDIA hardware only. If you need to run on AMD or Apple Silicon, this is the wrong stack and you are looking at SYCL, ROCm, or Metal instead.