NEWS · CODING · #489
TileGym agent translates 24 CUDA tile kernels from Python/Triton to cuTile Rust with near-parity performance
An agentic AI skill in the TileGym repository translates all 24 public TileGym operators (≈40 GPU kernels) from cuTile Python and Triton-TileIR into cuTile Rust (cutile-rs). The multi-agent, machine-checkable pipeline uses Tile IR diffs and staged validators and achieves a geomean parity of 0.995 versus cuTile Python (individual operators clear a 0.95 speedup threshold) on an NVIDIA DGX B200, while integrating translated kernels via a zero-copy C-ABI tensor descriptor layer.
KEY POINTS
- An agentic AI skill in the TileGym repository translates all 24 public TileGym operators (≈40 GPU kernels) from cuTile Python and Triton-TileIR into cuTile Rust (cutile-rs).
- The multi-agent, machine-checkable pipeline uses Tile IR diffs and staged validators and achieves a geomean parity of 0.995 versus cuTile Python (individual operators clear a 0.95 speedup threshold) on an NVIDIA DGX B200, while integrating translated kernels via a zero-copy C-ABI tensor descriptor layer.
- Demonstrates that an agentic, verifiable multi-agent workflow can port low-level GPU kernels into safe, idiomatic Rust with near-native performance, lowering friction for Rust-based GPU development and maintainable kernel migration.
WHY IT MATTERS
Demonstrates that an agentic, verifiable multi-agent workflow can port low-level GPU kernels into safe, idiomatic Rust with near-native performance, lowering friction for Rust-based GPU development and maintainable kernel migration.