Along with today’s NVIDIA R580 Linux driver beta, the CUDA 13.0 toolkit is now available to download and depends upon the new R580 Linux driver series.
The GA release of the CUDA Toolkit 13.0 is now available and pairs with the new R580 driver series. CUDA 13.0 supports Turing through Blackwell GPUs with older GPUs now dropped.
With CUDA 13.0, Arm platform support is now unified in the CUDA toolkit so that there is a single-install and consistent build across all Arm architectures (except for Jetson Orin based hardware, but moving forward they are pledging for this unified environment). The CUDA 13.0 release notes don’t outline anything net yet around NVIDIA’s plans for CUDA on RISC-V.
CUDA 13.0 also updates vector types with 32-bit alignment for greater load/store performance on Blackwell, support for newer enterprise Linux distributions, improved error reporting, GPU migration support with cuda-checkpoint, Fatbin file compression switched from LZ4 to Zstd, the CUDA runtime now uses context-less loading, and there are a variety of new CUDA APIs. CUDA 13.0 also adds host compiler support for LLVM Clang 20 and GCC 15.
CUDA 13.0 can be downloaded from developer.nvidia.com. The release notes have more details on the many new improvements and other changes with the CUDA 13.0 toolkit release.
