Nvidia brings CUDA Python 1.0 to C++ level

A single command is supposed to install the new Python environment: pip install cuda-python cuda-cccl numba-cuda-mlir[cu13]. Nvidia is releasing CUDA Python 1.0 alongside CUDA 13.3 and positioning the interface at the same level as C++.

The core is called cuda.core. It provides Python programs with objects for CUDA devices, streams — execution flows — and memory. This allows developers to write their own kernels, small programs for parallel computations, or optimize existing applications.

The package bundles several components: cuda.compute provides parallel algorithms such as sorting, searching and histograms, while cuda.bindings connects Python directly to all C interfaces. cuda-pathfinder locates installed CUDA components, and nvmath-python brings mathematical libraries to Python. According to Nvidia, this could make previously used packages such as PyTorch, CuPy and RAPIDS unnecessary.

According to Nvidia, the technical difference lies in the shared object model. Several components and applications access the same underlying CUDA objects, which can therefore be managed across library boundaries. This is also supposed to apply to Green Contexts, meaning kernels with particular latency advantages.

What happens next? According to Nvidia, Python developers get a more direct route to CUDA without necessarily having to use additional intermediary libraries to access the C++ functions. Nvidia promises a stable API with semantic versioning: incompatible changes are supposed to appear only in major versions, while functions will be removed according to predictable rules. Whether the promised unification will actually replace the existing libraries in day-to-day use will depend on adoption of the released environment.