Nvidia releases CUDA 13.5 with Vera Rubin general availability support

Nvidia has released CUDA 13.5 and the 620.70.06 Linux driver. The update transitions support for the Vera Rubin platform from preliminary testing to full production release.

Image: Phoronix

The release includes changes to how Windows handles compute cards. Nvidia notes that the Tesla Compute Cluster mode is incompatible with Rubin and subsequent GPU generations. Windows users must now use WDDM or Microsoft's Compute Driver Model (MCDM). Additionally, Windows on Arm support, which became available in CUDA 13.4, now extends to DGX Station for Windows.

CUDA 13.5 introduces several updates to GPU resource management and multi-GPU communication. The Multi-Process Service (MPS) received a third revision that adds weighted timeslicing and load balancing between SM partitions. It also allows administrators to route clients into namespaces through cgroup v2. For CUDA Tile, kernels can now use unicast pointers over NVLink to access memory on other GPUs. On NVSwitch systems, Tile kernels can employ multicast pointers alongside specialized multimem operations to handle data reductions and storage across the fabric.

Nvidia reports various performance improvements for its math libraries. On Blackwell hardware, cuBLAS provides up to 1.5x faster FP64 SYRK and HERK operations. cuSPARSE shows a 1.65x to 2.27x speedup for the ALG3 CSR SpMM path and a 2.79x geometric mean improvement for the sparse triangular solve analysis phase. nvJPEG hybrid GPU decode gains range from 10% to 600% depending on the workload.

While Tile-SIMT interoperability is in public preview, the ability to call a user's own device functions is still a planned feature.