NVIDIA CUDA Toolkit 13.0 Update 1 Brings Some Performance Enhancements
Released just over one month ago was the general availability of CUDA 13.0 while out this week is CUDA 13.0 Update 1 as the first incremental step forward to CUDA 13.
NVIDIA CUDA 13.0 Update 1 is now available with various fixes and other minor enhancements for those using this NVIDIA GPU compute stack on Windows and Linux.
Beyond routine fixes, CUDA 13.0 Update 1 does bring performance improvements for block-scaled FP4 GEMMs on NVIDIA Blackwell and Blackwell Ultra GPUs. There is also better performance for SYMV on Blackwell, cublasMatmul improvements for small cases, better TF32 GEMM performance on Thor GPUs, and kernel launch latency improvements. To the cuSPARSE library is also now support for the BSR format in the SpMV API. The rest of the NVIDIA CUDA 13.0 Update 1 changes appear mostly minor.
CUDA 13.0 Update 1 can be downloaded from developer.nvidia.com. More details on the NVIDIA CUDA 13.0 Update 1 changes via the release notes.
NVIDIA CUDA 13.0 Update 1 is now available with various fixes and other minor enhancements for those using this NVIDIA GPU compute stack on Windows and Linux.
👁 NVIDIA RTX 50 Blackwell GPUs
Beyond routine fixes, CUDA 13.0 Update 1 does bring performance improvements for block-scaled FP4 GEMMs on NVIDIA Blackwell and Blackwell Ultra GPUs. There is also better performance for SYMV on Blackwell, cublasMatmul improvements for small cases, better TF32 GEMM performance on Thor GPUs, and kernel launch latency improvements. To the cuSPARSE library is also now support for the BSR format in the SpMV API. The rest of the NVIDIA CUDA 13.0 Update 1 changes appear mostly minor.
CUDA 13.0 Update 1 can be downloaded from developer.nvidia.com. More details on the NVIDIA CUDA 13.0 Update 1 changes via the release notes.