The Future Is Tiled: Using CuTile & TileIR To Write Portable, High-performance GPU...
summary
NVIDIA introduces a new stack: cuTile Python DSL atop a virtual ISA called TileIR exposed as a TileIR MLIR dialect, with a TileIR backend for Torch Inductor and plans to ship TileIR in CUDA 13.1 and cuTile as a pip wheel. The system targets tensor cores and TMA, abstracts tiling/placement choices from users, enforces static tile shapes (currently power-of-two), and reports beta B200 results within 10% of hand-tuned state‑of‑the‑art kernels; cuTile is open source.
AI-written summary · about
chapters (12)
- 0:00Tiled GPU programming intro
- 2:48Evolving GPU programming model
- 5:15Introducing CUDA Tile
- 7:00TileIR virtual instruction set
- 9:47Programming model abstraction
- 12:57CuTile Python kernels
- 16:29Interoperability with CUDA
- 17:55Matrix multiplication in CuTile
- 20:43Performance results
- 21:05TileIR & CuTile roadmap
- 23:53Resources & future
- 24:35Q&A
related talks
01h 00mThe Shape of Things to Come
01h 26mGPU Algorithm Authoring with CUDA Tile
31mBuilding Inference Workflows with Tile Languages
22mTileLoom: Automatic Dataflow Planning for Tile-Based Languages on Spatial Dataflow Accel…
16mLightning Talk: Triton Compiler
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
14mWelder: Scheduling Deep Learning Memory Access via Tile-graph