News

NVIDIA releases CUDA 13.1 with a new tile-based programming model

Friday, December 5, 2025 at 11:29 PM

NVIDIA has launched CUDA 13.1, which the company describes as its largest update to the platform in 20 years, featuring the new CUDA Tile programming model for higher-level GPU algorithm development.

Context

NVIDIA released CUDA Toolkit 13.1 on December 4, 2025, a significant update described as the most comprehensive to the CUDA platform in two decades. The core innovation is CUDA Tile, a new tile-based programming model that includes a Virtual ISA (CUDA Tile IR) and a Python DSL (cuTile). This model elevates the abstraction level, enabling developers to define computations on data "tiles" while the system efficiently manages low-level hardware mapping, including tensor cores. This advancement is crucial for improving code portability across future GPU architectures and is initially designed for NVIDIA Blackwell GPUs (compute capabilities 10.x and 12.x). By abstracting intricate hardware details, CUDA Tile simplifies GPU programming, allowing developers to focus more on algorithms. Utilizing cuTile Python requires CUDA Toolkit 13.1, an NVIDIA driver R580 or later, and Python 3.10+. This strategic update reinforces NVIDIA's leadership in optimizing performance and developer experience for its advanced AI and computing platforms.

Related Companies

Nvidia
Nvidia
NVDA
US