CUDA Tile IR

license

Apache-2.0-WITH-LLVM-exception(OSI)

source

public

core features withheld

no

Apache-2.0-WITH-LLVM-exception license body confirmed. The repository is public and unarchived and builds the whole product, and the README describes no paid tier, enterprise edition or license-gated build beside it, so source is public and the core ungated.

• https://api.github.com/repos/NVIDIA/cuda-tile recorded 2026-09-02

  Repo metadata for cuda-tile's canonical repository - private false, archived false, fork false.

• https://github.com/NVIDIA/cuda-tile/blob/main/LICENSE.txt recorded 2026-10-07

  LICENSE body states the project is under the Apache License v2.0 with LLVM Exceptions and then runs the verbatim Apache-2.0 text plus the LLVM exception clause - real open source, not a closed-binary wrapper repository.

• https://github.com/NVIDIA/cuda-tile/blob/main/README.md recorded 2026-10-07

  README describes CUDA Tile IR as "an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization", made up of a CUDA Tile dialect, Python bindings, a bytecode format and a conformance test suite, in an open-source release "aligned with the **CUDA Toolkit 13.1** release"; it names no paid tier, enterprise edition or license-gated build beside the published source.

GitHub stars are the adoption signal for CUDA Tile IR, which the README says to build from source with CMake against an LLVM install rather than install as a package. The `cuda-tile` name on PyPI belongs to cuTile Python, a separate package from the nvidia/cutile-python repository pulled in by the CUDA Python stack, so its download count measures that different project rather than this compiler.

• https://api.github.com/repos/NVIDIA/cuda-tile recorded 2026-09-02

  Repo metadata - stargazers_count = 1,021 - for NVIDIA/cuda-tile.

CUDA Tile IR is a full compiler infrastructure - an MLIR dialect, Python bindings, a bytecode format and a conformance suite - for lowering tile-based tensor computations, narrowed to one vendor's targets, NVIDIA GPUs. That single-vendor scope is what keeps it short of Apache TVM's multi-vendor retargetability.

• https://github.com/NVIDIA/cuda-tile/blob/main/README.md recorded 2026-10-07

  README describes CUDA Tile IR as "an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization", made up of a CUDA Tile dialect, Python bindings, a bytecode format and a conformance test suite, in an open-source release "aligned with the **CUDA Toolkit 13.1** release"; it names no paid tier, enterprise edition or license-gated build beside the published source.