Develop High-Performance GPU Kernels in C++ with NVIDIA CUDA Tile - NVIDIA Developer

NVIDIA has introduced support for C++ in its CUDA Tile programming model with the release of CUDA 13.3, enabling developers to write highly optimized GPU kernels using a tile-based abstraction within ... Read original →

NVIDIA’s extension of CUDA Tile to C++ is less about developer convenience and more about tightening its software moat. Technically, it forces the entire HPC and AI stack—compilers, libraries, middleware—to align with NVIDIA’s tiled abstraction, raising integration costs for rivals like AMD ROCm and Intel oneAPI. Geopolitically, Chinese AI firms, already constrained by U.S. chip export controls, now face deeper dependency on a closed ecosystem, amplifying supply chain fragility. In response, AMD may double down on open-source collaboration, while Intel could push unified CPU-GPU compilation—but neither can quickly erode CUDA’s dominance. Within 18 months, as Blackwell and beyond roll out, tile-based kernel design will become the de facto standard, locking global developers into NVIDIA’s hardware-software-talent flywheel.