Fearless Concurrency on the GPU
Talkat RustConf 2026 · on Wednesday, 2026-09-09· at 11:55· Room 511 | Palais des Congrès de Montréal, Palais des Congrès de Montréal·
summary
NVIDIA presents a tile programming model implemented as Coutile Rust that JIT-compiles kernels to an open Tile IR and claims data-race-free GPU code with a formal data-race freedom proof. Host-side partitioning maps disjoint tensor tiles to device programs; most of Grout (the Hugging Face inference engine) kernels are safe, a few remain unsafe, and performance on a B200 nears cuBLAS for GEM and matches hand-tuned attention after bounds-check hoisting and JIT optimizations. Project is Apache 2.0 open source.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:06GPU programming risks
- 3:14Rust on the GPU today
- 5:25Tile programming model
- 8:15Coutile Rust overview
- 12:11Tile kernel basics
- 15:43Compilation process
- 17:07Safety guarantees
- 18:11Unsafe code in practice
- 20:43Bounds check optimizations
- 24:16Performance results
- 26:03Kernel composition
- 28:14Async GPU execution
- 31:52Future work and open source
- 34:24Q&A
related talks
30mKuiper: verified and efficient GPU programming
25m"Rust for AI & Accelerated Computing"
59mParallel Programming in Rust: Techniques for Blazing Speed
28mInterview
01h 22mCostless Software Abstractions for Parallel Architectures
01h 00mC++ on GPUs done right?
01h 07mWriting Data Parallel Algorithms on GPUs
26mBuilding Private And Secure Applications Easily Using The Sunscreen FHE Compiler&Rust