This email is to notify you of the release of NVIDIA CUDA Toolkit 13.1. This release introduces significant new features and performance enhancements across the CUDA programming model, software stack, developer tools, and math libraries. Users should review the changes to plan for integration and utilization of the new features.
New Features
- CUDA Tile: New tile-based programming model for GPUs.
- New CUDA programming guide.
- Green Contexts: Now available in the Runtime API for resource partitioning.
- CCCL Updates: New deterministic floating-point options for cub::DeviceReduce and simplified single-phase CUB APIs.
Developer Tools Updates:
- Added Nsight Compute support for profiling CUDA Tile kernels.
- Compute Sanitizer supports NVCC compile-time patching for memory checks.
Math Libraries & Performance Updates:
- cuBLAS adds Grouped GEMM support for Blackwell (FP8/BF16/FP16).
- Performance enhancements for cuBLAS and cuSOLVER on Blackwell architectures.
For a complete overview of all updates, features, compatibility details, and known issues, please review the official release notes documentation or read the CUDA Toolkit 13.1 blog.
