NVIDIA has introduced cuTile Rust, a tile-based system for authoring GPU kernels in the Rust programming language. This system extends Rust's ownership model to tile-based GPU kernels, enabling safe and idiomatic development. To facilitate the adoption of cuTile Rust, NVIDIA's team built an AI agent skill capable of translating existing cuTile Python and Triton-TileIR kernels.
This AI agent skill has been used to port all 24 public TileGym operators to cuTile Rust, achieving 99.5% of cuTile Python performance on average. The translation process employs a bounded multi-agent workflow that includes machine-checkable verdicts at each stage, such as IR diffing against reference Tile IR to verify structural equivalence before functional testing.
Key Points
- cuTile Rust is a tile-based system for safe, idiomatic GPU kernel authoring in Rust.
- An AI agent skill translates cuTile Python and Triton-TileIR kernels into cuTile Rust.
- All 24 public TileGym operators have been ported to cuTile Rust using this skill.
- Translated kernels achieve 99.5% of cuTile Python performance on average.
- The translation pipeline uses a bounded multi-agent workflow with machine-checkable verdicts at each stage.
- IR diffing against reference Tile IR verifies structural equivalence before functional testing.
- cuTile Rust requires explicit specialization in kernel signatures, unlike cuTile Python's implicit specialization at call time.
Context
According to the NVIDIA Developer Blog, cuTile Python, Triton-TileIR, and cuTile Rust are front ends over the same CUDA Tile IR (cuda_tile dialect). All three feed into the same tileiras compiler, which performs tile-level optimizations and emits the GPU binary. This shared foundation makes translation across the CUDA Tile family practical and verifiable.
Why It Matters
This development provides a pathway for developers to leverage the safety and ownership model of Rust for GPU kernel programming, while maintaining performance parity with existing cuTile Python implementations. The use of agentic AI for translation and machine-checkable verdicts enhances the reliability and efficiency of porting efforts.
What To Do
- Access the tilegym-converting-python-to-rust agent skill to translate your own kernels.
- Explore the converted cuTile Rust operators in the TileGym repository.
- Review the TileGym README for setup requirements and usage instructions.