NVIDIA Releases CUDA Toolkit 13.5 with Vera Rubin Support

NVIDIA released CUDA Toolkit 13.5, introducing general availability (GA) support for the next-generation Vera Rubin architecture and major upgrades to multi-GPU communication and MPS v3 scheduling. The update also brings optimized low-precision math libraries and a new Nsight Perf SDK for AI. This release transitions NVIDIA's Vera Rubin AI platform from developer preview to enterprise-grade production, enabling high-performance training and inference. The enhancements significantly boost multi-GPU efficiency and resource allocation for large-scale data center deployments. CUDA 13.5 introduces direct cross-GPU memory access via CUDA Tile and NVLink, alongside MPS v3 features like weighted time slices and Streaming Multiprocessor (SM) load balancing. Additionally, Windows support drops the legacy TCC driver model in favor of WDDM and MCDM.