CUDA to cuTile transpiler for Nvidia's CUDA 13.1 (github.com)

NVIDIA has introduced the RightNow Tile, a transpiler that converts traditional CUDA SIMT (Single Instruction, Multiple Threads) kernels into cuTile Python code, specifically optimized for the new Blackwell GPUs. This transpilation represents a significant shift in GPU programming from thread-centric to tile-centric paradigms, allowing for automatic memory management and simpler code semantics. RightNow Tile is part of a broader ecosystem designed to enhance GPU kernel development, making it easier for developers to adapt their existing CUDA applications to leverage the performance advantages of next-generation hardware. The tool automatically analyzes CUDA kernels, identifying 18 computational patterns and applying over 60 optimizations tailored to specific GPU tasks. It handles various complex operations like matrix multiplications and attention mechanisms used in deep learning, significantly reducing the complexity of manual optimization. With comprehensive support for different computational models and an intuitive interface built with modern web technologies, RightNow Tile empowers AI/ML developers to maximize performance on Blackwell architecture while streamlining the programming process.