NVIDIA releases CUDA 13.5 with Vera Rubin general availability

NVIDIA CUDA 13.5 reaches general availability with production-ready support for the Vera Rubin GPU architecture. Windows on Arm support, which has been GA since CUDA 13.4, now extends to NVIDIA DGX Station for Windows. Core math libraries also gain performance work on Vera Rubin, including faster NVFP4 and MXFP8 matrix multiplication.

NVIDIA Multi-Process Service v3 adds weighted timeslicing, SM partition load balancing with borrowing and lending, cgroup-based routing, and per-namespace SM partition assignments. CUDA Graphs can now capture existing cudaMemcpyBatchAsync calls. Tesla Compute Cluster mode is not supported on Rubin and later GPUs.

The cuda.core 1.2 Python library adds copy_batch() for batched memory copies and Programmatic Dependent Launch through LaunchConfig. CCCL 3.5 adds public CUB tuning APIs and a batched Top-K selection that is integrated into FlashInfer. A public preview lets CUDA Tile kernels call device functions from libraries such as NVSHMEM and Warp.

Teams running agent inference or training on Vera Rubin hardware can now treat that support as production-ready. Deployment tools that assume TCC mode should switch to the WDDM or MCDM configuration on Rubin systems.

Developers writing GPU code in Python can batch copies and update graph parameters without reinstantiating graphs, which helps when inputs change between runs. The batched Top-K selection is relevant to token selection in LLM inference.