Nvidia's CUDA Tile examined: AI giant releases programming style for Rubin, Feynman, and beyond — tensor-native execution model lays the foundation for Blackwell and beyond

Nvidia's latest CUDA 13.1 release introduces the CUDA Tile programming path, shifting from SIMT to a tensor-native execution model, laying the groundwork for future architectures like Blackwell. This change allows developers to focus on data operations rather than low-level hardware details, improving performance scalability across GPU generations. The CUDA Tile IR serves as a virtual instruction set for tile workloads, while cuTile Python enables developers to create array- and tile-oriented kernels directly in Python. This shift aims to optimize performance for AI-centric algorithms initially and expand to various applications like scientific simulations and HPC workloads in the future.