NVIDIA has announced CUDA Toolkit 13.2, featuring CUDA Tile, a new tile-based GPU programming model that simplifies...

NVIDIA has announced CUDA Toolkit 13.2, featuring CUDA Tile, a new tile-based GPU programming model that simplifies kernel development by abstracting hardware specifics. CUDA Tile now supports Ampere and Ada architectures (previously only Hopper and newer) and introduces recursion and closures in cuTile Python. This expansion improves programmability for older GPUs, potentially broadening the developer base and accelerating deployment from data center to edge. The official developer blog (https://developer.nvidia.com/blog/cuda-13-2-introduces-enhanced-cuda-tile-support-and-new-python-features/) details these enhancements. Separately, a prediction market on Manifold (https://manifold.markets/_deleted_/will-cuda-remain-a-monopoly-for-gpu) as of the same date shows a 56.95% probability that CUDA will maintain its monopoly through 2027, indicating modest market confidence in NVIDIA's software lead. Taken together, the release strengthens NVIDIA's competitive position by making CUDA more accessible and feature-rich, potentially increasing developer stickiness and slowing the adoption of alternatives like AMD ROCm or Intel oneAPI. The move aligns with NVIDIA's strategy to defend its AI infrastructure moat through continuous software innovation.