A new mechanism for fine-grained GPU resource partitioning, allowing applications to dynamically allocate Streaming Multiprocessors (SMs) for deterministic workload performance. 🛠️ Performance & Technical Updates
cuBLAS and cuFFT have introduced new opaque API structures in CUDA 13. Code compiled against CUDA 12 libraries will need recompilation. NVIDIA provides a cuda-compat-13 package for RHEL/Ubuntu to ease transition, but mixed-version linking is discouraged. cuda toolkit update news
: Features like 32-byte alignment for vector types and improved fatbin compression (using Zstandard) mean faster execution and smaller binary sizes. A new mechanism for fine-grained GPU resource partitioning,
The headline feature of 12.9 is official, production-grade support for the . Key updates include: cuda toolkit update news
