CUDA 13.5 moves Vera Rubin support to general availability
NVIDIA released CUDA Toolkit 13.5 on November 4, 2026 with production-ready Vera Rubin GPU architecture support, plus GPU scheduling updates, multi-GPU CUDA Tile, CUDA Python, CCCL, and Nsight tooling changes.
Why it matters
Per NVIDIA’s technical blog, Vera Rubin support moves from early developer access to general availability framed as ready for enterprise production. Windows on Arm support, GA since CUDA 13.4, now extends to NVIDIA DGX Station for Windows. From the Rubin architecture onward, CUDA supports WDDM and MCDM; TCC mode is not supported on Rubin and later GPUs.
Locality domains and Multi-Process Service v3 add scheduling control: weighted timeslicing, SM partition load balancing with borrow/lend, cgroup-based routing, and per-namespace partition assignments. CUDA Tile adds multi-GPU communication via unicast and multicast pointers on NVLink and NVSwitch, plus a public preview of Tile–SIMT interoperability so Tile kernels can call device functions from libraries such as NVSHMEM and CCCL.
CUDA Python cuda.core 1.2 adds batched memory copies, Programmatic Dependent Launch, and CUDA Graph updates without reinstantiation. CCCL 3.5 brings public CUB tuning APIs, reproducible floating-point scans, and batched Top-K. The toolkit and release notes are available on NVIDIA’s developer site.