• NVIDIA CUDA Tile is a GPU-based programming model that targets portability for NVIDIA Tensor Cores, unlocking peak GPU performance. • One of the great things about CUDA Tile is that you can build your own DSL on top of it. • This post shares the work NVIDIA is doing to integrate CUDA Tile as a backend for OpenAI Triton, an open source Python DSL designed to write DL kernels for GPUs. • OpenAI Triton supports tiled computation, a technique that divides data and computational tasks into small blocks. • Triton contains an MLIR-based compiler that generates PTX. • This enables researchers without CUDA experience to write efficient GPU code.

Article Summaries:

  • NVIDIA has added a CUDA Tile IR backend to OpenAI Triton, an open‑source Python DSL for writing GPU kernels. The new Triton‑to‑TileIR bridge lets developers compile Triton code directly to CUDA Tile IR-an MLIR‑based representation that natively supports tile‑level computation on NVIDIA Tensor Cores-rather than the traditional PTX backend. This integration preserves Triton’s tile‑based abstractions, automates thread scheduling and resource allocation, and offers a single‑environment‑variable switch to choose between PTX and TileIR per kernel. The move aims to simplify GPU programming, improve performance, and enhance portability across NVIDIA’s next‑generation GPU architectures.

Sources: