Nvidia's CUDA 13.5 Moves Vera Rubin Support to General Availability

CUDA is the programming platform that most AI software running on Nvidia hardware is built on, so general availability is the point at which developers and cloud operators can treat Rubin as a supported target for production workloads, not just for testing. The previous version, CUDA 13.4, gave developers early access to Rubin features and added Windows on Arm support.

Vera Rubin is the successor to Nvidia's Blackwell generation, and the chips are already the basis of large financing deals. This week Goldman Sachs and Apollo began offering investors pieces of a $40 billion lease financing for SpaceX's Vera Rubin chips.

The post, by Jonathan Bentz, who leads Nvidia's CUDA technical marketing engineering team, lists changes across the software stack. In the cuBLAS maths library, matrix multiplications in the low-precision NVFP4 and MXFP8 formats, which are widely used to train and run AI models, now run faster on Rubin. On Rubin and later GPUs, CUDA supports the WDDM and MCDM driver modes on Windows but not TCC.

Several of the changes are aimed at sharing large GPUs between workloads. The third version of Nvidia's Multi-Process Service adds weighted timeslicing. It also lets partitions of streaming multiprocessors borrow and lend capacity to balance load, and can route work to separate namespaces using Linux cgroup v2. Applications can also now split a GPU into partitions of different sizes.

CUDA Tile, Nvidia's newer tile-based programming model, gains multi-GPU communication: kernels can reach memory on GPUs connected by NVLink and run reductions across NVSwitch fabrics. A public preview lets tile kernels call functions from Nvidia libraries written in the traditional SIMT style. Nvidia said support for developers' own device functions would come in a later release.

On the tools side, Nvidia released a public preview of the Nsight Perf SDK for AI, which is meant for always-on profiling of long-running workloads. It also added a Compute Sanitizer skill for AI coding agents and OpenTelemetry trace integration in Nsight Cloud. The CCCL 3.5 libraries add a batched top-k operation, which Nvidia said is integrated into the FlashInfer inference library, and an option to make floating-point scans give the same result on every run.

Windows on Arm support, generally available since CUDA 13.4, now extends to the DGX Station for Windows. The cuSPARSE and nvJPEG libraries also get performance gains: Nvidia said some sparse matrix operations run between 1.65 and 2.27 times faster, and some JPEG decode throughput improves by 10% to 600% depending on the image.

Developers whose scripts parse the output of the cuobjdump tool may need to update them, because the tool now prints more descriptive symbol-table columns. CUDA Toolkit 13.5 is available to download from Nvidia's CUDA downloads page, and the release notes give full platform and compatibility details.