NVIDIA/cutile-python

cuTile is a programming model for writing parallel kernels for NVIDIA GPUs

What it solves

cuTile Python provides a way to write GPU kernels using a Python-based programming language, removing the need to write low-level CUDA C++ for parallel operations on NVIDIA GPUs.

How it works

It generates kernels based on Tile IR (Intermediate Representation). It allows developers to define kernels using Python decorators (like @ct.kernel) and manage data movement between memory and GPU tiles using load and store operations. These kernels are then launched on the GPU via a grid configuration.

Who it’s for

Developers who want to program NVIDIA GPUs using Python instead of C++, specifically those targeting Blackwell, Ampere, or Ada architectures.

Highlights

  • Pythonic GPU Programming: Write parallel kernels directly in Python.
  • Tile-based Memory Management: Uses explicit load and store operations to handle data tiles.
  • Broad Integration: Works with libraries like CuPy and supports DLPack.
  • Hardware Support: Specifically optimized for NVIDIA Blackwell, Ampere, and Ada GPUs.

Related

  • Project
    NVIDIA/tilus

    Tilus is a tile-level GPU kernel programming language that enables high-performance, low-precision computation with explicit control over shared memory and registers.

  • Project
    NVIDIA/cuda-tile

    CUDA Tile IR is an MLIR‑based compiler infrastructure from NVIDIA that introduces a dedicated dialect for tile‑oriented CUDA kernels. It provides C++ and Python APIs, a binary bytecode format, and tools to translate MLIR to bytecode or cubin, enabling high‑performance kernels that target NVIDIA Tensor‑Core units. The project builds on LLVM/MLIR, aligns with CUDA Toolkit 13.1, and includes a conformance test suite. It is a genuine AI‑related compiler project, not a tutorial or list.

  • Project
    NVlabs/cutile-rs

    A tile-based system for writing memory-safe, data-race-free GPU kernels in Rust using JIT compilation via CUDA Tile IR.

  • Project
    NVIDIA/cuda-python

    A collection of Python packages providing both high-level idiomatic and low-level binding access to NVIDIA's CUDA platform for GPU-accelerated computing.

  • Project
    NVIDIA/TileGym

    A CUDA Tile kernel library providing tutorials and examples for tile-based GPU programming to optimize deep learning operators and LLM inference.