NVIDIA has released CUDA Toolkit 13.5 with general availability support for the NVIDIA Vera Rubin GPU architecture. The release also extends Windows on Arm support to NVIDIA DGX Station for Windows. New controls include locality-domain SM partitioning, weighted timeslicing and load balancing in NVIDIA Multi-Process Service v3, and batched memory copies in CUDA graphs. CUDA Tile adds multi-GPU communication and a public preview of interoperability with SIMT library device functions. CUDA Toolkit 13.5 is available for download, with runtime API and MPS documentation covering the new features.
Why it matters
The release gives developers production-ready support for Vera Rubin and more control over how GPU resources are divided among workloads. Its library, graph, Python and multi-GPU updates can help builders tune resource sharing and integrate accelerated operations into existing applications.
Key points
- CUDA Toolkit 13.5 provides general availability support for the NVIDIA Vera Rubin architecture.
- Windows on Arm support now extends to NVIDIA DGX Station for Windows.
- MPS v3 adds weighted timeslicing, SM partition load balancing and cgroup-based routing.
- CUDA Tile supports multi-GPU communication over NVLink and NVSwitch fabrics.
- CUDA Python cuda.core 1.2 adds batched copies, Programmatic Dependent Launch and flexible CUDA Graph updates.
Based on reporting by NVIDIA Technical Blog.
Join the conversation! Your thoughts help the community grow.