NVIDIA CUDA 13.5 reaches general availability with Vera Rubin GPU support
CUDA 13.5 introduces production-ready support for the Vera Rubin architecture alongside significant scheduling and memory management improvements via MPS v3.
Summary
NVIDIA has released CUDA Toolkit 13.5, marking the transition of the Vera Rubin GPU architecture from early developer access to general availability (GA). This update is designed for enterprise production environments, providing the necessary driver and toolkit support for the next generation of NVIDIA hardware.
The release also introduces substantial enhancements to the NVIDIA Multi-Process Service (MPS) and locality domain management. These changes allow for more granular control over how GPU resources, such as Streaming Multiprocessors (SMs), are partitioned and shared among multiple workloads or processes.
Technical advancements in CUDA 13.5
- Vera Rubin Architecture: Moves to general availability (GA), enabling enterprise-scale production deployment on new NVIDIA hardware.
- MPS v3 Scheduling: Introduces weighted timeslicing for prioritized compute turns and SM partition load balancing, which allows partitions to borrow or lend unused SM capacity.
- Locality Domain Management: Adds the
cudaDevSmResourceSplit()API and new flags to allow applications to choose between uniform SM partitions or maximizing hardware utilization by including all available SMs in a domain. - CUDA Tile & Communication: Enables multi-GPU communication via unicast and multicast pointers on NVLink/NVSwitch, alongside a public preview of SIMT interoperability with libraries like NVSHMEM and CCCL Coop.
- Memory & Graph Operations: Adds support for batched 1D memory copies within CUDA Graphs and introduces polling-based memory accounting in MPS as a fallback for unreliable cgroup notifications.
Why it matters
The move to GA for Vera Rubin ensures that enterprise AI and high-performance computing (HPC) developers can transition from testing to production on NVIDIA's latest silicon. Furthermore, the improved MPS scheduling and locality domain controls provide system administrators and cloud providers with better tools to optimize GPU density and resource isolation in multi-tenant environments.
What to do
- Download CUDA Toolkit 13.5 to use the new features.
- Review the CUDA Runtime API documentation for locality domain partition details.
- Explore the MPS v3 interface guide to configure weighted timeslicing and SM partition load balancing.
- Run nvidia-cuda-mps-control memacct set force_polling=true before memory accounting starts to use polling-based memory accounting as a fallback.
Original reference: developer.nvidia.com →