cuda.bindings

Type: Technology Tags: NVIDIA, CUDA, Python, Bindings, Driver API, Runtime API, NVRTC, NVML Related: CUDA-Python, cuda-core, CUDA-Driver-API, CUDA-Runtime-API, NVRTC, PTX-ISA, Nsight-Compute, Nsight-Systems, CuPy Sources: https://nvidia.github.io/cuda-python/cuda-bindings/latest/overview.html, https://developer.nvidia.com/cuda/python Last Updated: 2026-04-29

Summary

cuda.bindings provides NVIDIA-maintained low-level Python bindings to CUDA C APIs. It is the API-parity layer of CUDA-Python, intended to give Python projects broad access to CUDA host APIs while improving portability and interoperability across the Python GPU ecosystem.

Detail

Purpose

Some Python libraries need direct control over CUDA objects, modules, memory, streams, events, compiled PTX, runtime APIs, or management APIs. cuda.bindings serves that need by exposing low-level CUDA APIs from Python rather than hiding them behind higher-level abstractions.

Key capabilities

• Low-level Python access to CUDA driver/runtime API families and related CUDA host APIs.
• Workflows for compiling device code to PTX, loading modules, extracting functions, preparing kernel arguments, launching kernels, and synchronizing streams.
• Compatibility with profiling workflows using Nsight-Systems and Nsight-Compute.
• CUDA native object wrappers that expose pointer-style access through getPtr() and integer conversion where needed.
• Foundation layer used by higher-level Python CUDA components such as cuda-core.

NVIDIA context

cuda.bindings is the closest Python counterpart to CUDA-Driver-API, CUDA-Runtime-API, and NVRTC C APIs. Use it in the wiki when a question is about exact CUDA host API access from Python; use cuda-core when a question is about Pythonic higher-level CUDA objects and workflows.

Connections

• CUDA-Python - umbrella that includes cuda.bindings.
• cuda-core - higher-level Pythonic CUDA layer built above low-level bindings.
• CUDA-Driver-API and CUDA-Runtime-API - core CUDA C APIs exposed through Python bindings.
• NVRTC - runtime compilation API commonly used before loading modules and launching kernels.
• PTX-ISA - compiled PTX may be loaded and launched through driver-style binding workflows.
• Nsight-Systems and Nsight-Compute - profiling tools compatible with CUDA Python binding examples.
• CuPy - example of the wider Python CUDA ecosystem that benefits from unified bindings.

Source Excerpts

• NVIDIA says cuda.bindings aims to unify the Python CUDA ecosystem with a standard set of low-level interfaces and full CUDA host API coverage.
• The documentation shows binding workflows for compiling PTX, loading CUDA modules, launching kernels, and preparing kernel arguments.