Translating CUDA Tile Operations from Python to Rust Using Agentic AI
Source
developer.nvidia.com
Author
Tanya Lenz
Date
Context
NVIDIA describes an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → workflow for translating GPU kernels from cuTile Python or Triton-TileIR into cuTile Rust. The key verification aid is that these front ends share a compiler representation. Comparing that representation exposes changes in operations, tile shapes or reductions that ordinary output tests might miss, as the technical article explains.
Separate stages analyze the reference, write the kernel, connect it to the host program and test correctness before benchmarking. Across 24 operators, the reported aggregate GPU execution performance was approximately equal to cuTile Python. That result measures kernel device time; application-level timing also includes launch and scheduling overhead.
Key takeaways · AI-distilled
NVIDIA reports that its AI-agent workflow translated all 24 public TileGym GPU operators into cuTile Rust, reaching 99.5% of the cuTile Python baseline's performance on average. The conversion was checked for both numerical correctness and speed.
Each translation stage ends in an automated check. Validator scripts and comparisons of the generated intermediate code against reference code decide whether the conversion advances, giving the agents concrete feedback rather than accepting plausible-looking output.
A core translation challenge is that cuTile Python specializes kernels implicitly at call time via JIT, while cuTile Rust requires every specialization (like tile size or dtype) to be declared explicitly in the kernel signature so rustc can check it at compile time.
Benchmarked on an NVIDIA DGX B200 across 347 paired configurations, all 24 converted operators cleared a 0.95 geomean speedup threshold versus the cuTile Python baseline, with an overall geomean of 0.995.
Using the agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition → for kernel conversion cut tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → cost to about half on average, while still validating every operator for numerical correctness and clearing the 0.95 geomean speedup bar.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
articleScaling NVIDIA TensorRT Performance Optimization Beyond a Single Build Route with the Global Performance Tuner, Part 2
articleAccelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tile
articleScale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
articleOptimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym
