CUDA 13.0 Update 2 is now available as the latest incremental improvement to NVIDIA's compute stack.

Following the [1]CUDA 13.0 GA in August and [2]CUDA 13.0 Update 1 in September, out today is CUDA 13.0 Update 2 with a few improvements for this Windows and Linux compute stack.

CUDA 13.0 Update 2 is still paired with the NVIDIA 580 series LInux driver, now version 580.95.05. CUDA 13.0 Update 2 does bring some performance improvements for the new NVIDIA DGX Spark hardware. Notable is improving the DGX Spark performance for FP16/BF16 and FP8 GEMMs with the cuBLAS library.

CUDA 13.0 Update 2's cuBLAS also now features opt-in fixed-point emulation for FP64 MATMULs to improve performance and power efficiency. This emulation follows the Ozaki-1 Scheme and uses an automatic dynamic precision framework for FP64-level accuracy.

This update also adds support to cuBLAS for BF16x9 FP32 emulation.

Downloads and more details on the CUDA 13.0 Update 2 release via [3]NVIDIA.com .



[1] https://www.phoronix.com/news/NVIDIA-CUDA-13.0

[2] https://www.phoronix.com/news/NVIDIA-CUDA-13.0-Update-1

[3] https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html