NVIDIA CUDA 13.5 reaches GA with Vera Rubin support and new developer features

NVIDIA has released CUDA Toolkit 13.5 to general availability, bringing production-ready support for the NVIDIA Vera Rubin GPU architecture. The release also extends Windows on Arm support to NVIDIA DGX Station for Windows and adds scheduling, communication, library, and tooling updates for developers.

Scheduling and MPS updates

CUDA 13.5 adds locality domain enhancements that let applications choose between uniform streaming multiprocessor partitions or partitions that include all available SMs. The cudaDevSmResourceGroupLocalityDomainSupplementalSms flag combined with cudaDevSmResourceGroupLocalityDomainId includes additional SMs within selected domains, and the cudaDevSmResourceSplit API can discover each partition SM count.

NVIDIA Multi-Process Service v3 introduces weighted timeslicing, SM partition load balancing with borrowing and lending, cgroup-based routing, per-namespace SM partition assignments, and a polling-based fallback for memory accounting. These features give more control over how workloads share a GPU.

CUDA Tile and CUDA Python

CUDA Tile gains multi-GPU communication with unicast and multicast pointers on NVLink and NVSwitch fabrics. A public preview of SIMT interoperability allows Tile kernels to call device functions from libraries such as NVSHMEM, CCCL Coop, Warp, and MathDx, enabling communication and computation inside a kernel.

CUDA Python cuda.core 1.2 adds batched memory copies through copy_batch, Programmatic Dependent Launch through LaunchConfig, and more flexible CUDA Graph updates without reinstantiation. These changes expand Pythonic access to CUDA programming.

CCCL and developer tools

CCCL 3.5 provides public tuning APIs for CUB, reproducible floating-point scans, batched Top-K selection integrated into FlashInfer, and data-dependent reductions that consume element counts from device memory using cuda::args::deferred. The release also includes the public preview of NVIDIA Nsight Perf SDK for AI, a timeline page in NVIDIA Nsight Compute, and CUDA-GDB support for stopping on CUDA Driver log events.

Core CUDA math libraries receive performance improvements on Vera Rubin, including faster GEMMs and new cuBLAS and cuSPARSE features. The update also adds compiler and Windows driver mode changes, with WDDM and MCDM supported on Rubin and later GPUs while TCC is not supported.

Sources: NVIDIA Developer Blog