Show HN: cuTile Rust: Safe, Data-Race-Free GPU Kernels Using Rust

Developed by NVIDIA Labs, cuTile Rust (cutile-rs) is a tile-based system that extends Rust's Ownership principles to the GPU runtime boundary, ensuring memory safety and supporting the writing of GPU kernels that are free of data-races.

The key features are as follows:

• Memory Safety: Mutable tensors are partitioned before execution to prevent data-races, while immutable tensors are shared for safe access.
• JIT Compilation: Via the #[cutile::module] macro, it captures the Rust AST and JIT-compiles it into a GPU cubin through CUDA Tile IR.
• High Performance: In tests on the NVIDIA B200, it recorded performance reaching 91% of peak memory bandwidth for element-wise operations and 92% of peak performance for GEMM operations.
• Flexibility: In addition to synchronous execution, it supports asynchronous pipelines and CUDA Graph Replay, and allows opt-out for low-level control when needed.

While currently in early development as a research project, it demonstrates GEMM performance competitive with cuBLAS, presenting new possibilities for GPU programming within the Rust ecosystem.