cuda.core

Type: Technology Tags: NVIDIA, CUDA, Python, Runtime, Streams, CUDA Graphs, Memory, JIT Related: CUDA-Python, cuda-bindings, CUDA-Runtime-API, CUDA-Driver-API, CUDA-Streams, CUDA-Graphs, NVRTC, nvJitLink, CuPy, PyTorch Sources: https://nvidia.github.io/cuda-python/cuda-core/latest/, https://nvidia.github.io/cuda-python/cuda-core/latest/getting-started.html, https://nvidia.github.io/cuda-python/cuda-core/latest/installation.html, https://nvidia.github.io/cuda-python/cuda-core/latest/interoperability.html Last Updated: 2026-04-29

Summary

cuda.core is NVIDIA’s Pythonic CUDA interface for core runtime-style functionality. It provides high-level Python constructs for devices, streams, events, memory resources, CUDA graphs, kernel compilation/launch, linking, Tensor Memory Accelerator descriptors, and CUDA/NVML system inspection.

Detail

Purpose

cuda.core is for Python developers who need direct CUDA control but do not want one-to-one CUDA C API bindings for every call. It sits above cuda-bindings by offering higher-level classes and workflows while still connecting to core CUDA concepts such as CUDA-Streams, CUDA-Graphs, memory management, and runtime code generation.

Key capabilities

• Compile and launch CUDA kernels from Python.
• Coordinate asynchronous execution with streams and events.
• Allocate, transfer, and manage device, pinned, managed, graph, and virtual memory resources.
• Build and execute CUDA graphs with GraphBuilder and graph-related classes.
• Runtime-link device code with link-time optimization flows.
• Inspect CUDA system state through driver/NVML-oriented APIs.
• Interoperate with other Python GPU libraries through current device/context sharing, a proposed __cuda_stream__ protocol, and memory-view utilities.

NVIDIA context

cuda.core is one of the most important current CUDA Python components because it offers a Python-native control surface for the same runtime concepts documented in CUDA-Runtime-API and CUDA-Driver-API. It is especially useful for Python packages that need CUDA streams, CUDA graphs, compiled kernels, or portable stream/memory exchange with CuPy, PyTorch, and other GPU array libraries.

Connections

• CUDA-Python - umbrella documentation that includes cuda.core.
• cuda-bindings - lower-level binding layer for one-to-one CUDA C APIs.
• CUDA-Runtime-API and CUDA-Driver-API - native CUDA API layers that cuda.core abstracts in Pythonic form.
• CUDA-Streams and CUDA-Graphs - core execution concepts surfaced by cuda.core.
• NVRTC and nvJitLink - related runtime compilation and linking concepts.
• CuPy and PyTorch - Python GPU libraries that can interoperate through streams, contexts, and memory views.

Source Excerpts

• NVIDIA describes cuda.core as Pythonic access to the CUDA runtime and other core functionality.
• NVIDIA’s cuda.core overview lists devices, memory resources, programs, graph builders, streams, and events as high-level constructs.