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, the CUDA Tile programming model, has triggered discussions in the industry about whether NVIDIA's "moat" will be weakened.Well-known chip architect Jim Keller believes this may mark the end of the software's exclusivity.
Kuai Technology News on December 9,
CUDA Tile is a model based on data blocks (tiles). Its design goal is to significantly lower the threshold for GPU programming. Developers can focus on organizing data into blocks and performing calculations, while the underlying complex thread scheduling, memory layout, and hardware resource mapping are automatically handled by the compiler and runtime system.
To support this new model, CUDA 13.1 introduced the virtual instruction set (Tile IR) and released the cuTile tool, allowing developers to use popular Python to write Tile-based GPU Kernel.
Jim Keller, a senior figure in the chip design industry who has participated in the design of AMD Zen, Apple A series and Tesla Autopilot, believes that if mainstream GPU programming in the future gradually shifts to this Tile-based approach, once developers get used to the "write Tile, hardware optimization" model.
Then the same set of program logic will be easier to transplant to different GPU hardware. It is not as highly bound to NVIDIA hardware as CUDA C++ in the past. This may provide entry opportunities for AMD, Intel or emerging AI companies.
