NVIDIA recently officially launched CUDA 13.1, which the official positioned as "the largest and most comprehensive upgrade since the birth of the CUDA platform in 2006."The core highlight of this update is the introduction of the revolutionary CUDA Tile programming model, marking the GPU programming paradigm entering a new, higher abstraction stage.

Traditional GPU programming is based on SIMT (Single Instruction Multi-Threading) mode, and developers need to pay attention to low-level details such as threads, memory, and synchronization.

CUDA Tile is a tile (tile, data block)-based model. Developers can now focus on organizing data into blocks and performing calculations on these data blocks.The underlying complex work of thread scheduling, memory layout, and hardware resource mapping will be automatically handled by the compiler and runtime.

To support Tile programming, CUDA 13.1 introduced the virtual instruction set (Tile IR) and released the cuTile tool to allow developers to use Python to write Tile-based GPU Kernel.

This greatly lowers the threshold for GPU programming, allowing data scientists and researchers who are not familiar with traditional CUDA C/C++ or underlying SIMT models to write GPU-accelerated code.

Tile programming does not replace SIMT, but provides a coexisting alternative path. Developers can flexibly choose the most appropriate programming model according to specific application scenarios.

The significance of CUDA 13.1 lies not only in adding new features or optimizing performance, but also in laying the foundation for building a new generation of high-level, cross-architecture GPU computing libraries and frameworks. By introducing Tile IR and high-level abstraction, NVIDIA has added a thicker middle layer between hardware and software.

In the past, competitors (such as AMD's ROCm and Intel's OneAPI) mainly relied on the compatibility layer for CUDA code translation, but for the new higher-abstraction model of CUDA Tile, simple code translation is no longer enough.

Competitors must build equally intelligent compilers to handle Tile IR, which undoubtedly increases the difficulty of technology alignment and objectively further increases the stickiness and user lock-in of the CUDA ecosystem.