As a wonderful Christmas gift to open-source fans, NVIDIA dropped their proprietary license on the CUDA Tile intermediate representation and has now made the IR open-source software.
Earlier this month with CUDA 13.1 NVIDIA announced the CUDA Tile programming model as “the largest and most comprehensive update to the CUDA platform since it was invented two decades ago” The intermediate representation (IR) to CUDA Tile is now open-source software.
The CUDA Tile IR is based on the MLIR intermediate representation out of the LLVM project. CUDA Tile IR being derived from Multi-Level Intermediate Representation (MLIR) is very significant and could help in seeing CUDA Tile IR being supported in non-NVIDIA environments. AMD makes use of MLIR to varying degrees in their AI/compute environment, Google’s IREE plays a big role here across different vendors, Intel has a dialect of MLIR for their hardware, etc. Various IR frameworks like ONNX-MLIR, MLIRE-AIE, Torch-MLIR, and others exist too. With CUDA Tile IR being based on MLIR could help in lowering it and ultimately bringing CUDA Tile IR targeting software to other GPUs/accelerators. At the very least CUDA Tile IR being open-source will eventually help efforts like ZLUDA.
The CUDA Tile code consists of the Tile MLIR dialect, Python API bindings, the bytecode representation, and a conformance test suite.
“CUDA Tile IR is an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization, focusing on tile-based computation patterns and optimizations targeting NVIDIA tensor core units. The project provides a comprehensive ecosystem for expressing and optimizing tiled computations for NVIDIA GPUs, simplifying the development of high-performance CUDA kernels through abstractions for common tiling patterns, memory hierarchy management, and GPU-specific optimizations.
It’s fabulous seeing NVIDIA make CUDA Tile IR open-source and will be interesting to see what the future holds in 2026. The open-source code can be found on GitHub. The open-source code is under an Apache 2.0 license.
