NVIDIA CUDA 13.5 Introduces Vera Rubin GA and GPU Scheduling Updates

NVIDIA has officially released CUDA 13.5, marking a significant milestone in GPU computing and artificial intelligence development. The update transitions NVIDIA Vera Rubin platform support from early developer access to general availability (GA), enabling enterprise production deployments. CUDA 13.5 also introduces enhanced GPU scheduling capabilities, expanded platform compatibility, and updates across its libraries and developer tools.

Vera Rubin Platform Now in General Availability

The Vera Rubin platform, which combines Rubin GPUs and Vera CPUs via NVLink-C2C, is now production-ready with CUDA 13.5. This follows NVIDIA’s June 2026 announcement of Vera Rubin systems featuring up to 144 GPUs per rack and delivering over 7 exaflops of AI-for-science performance and 5 petaflops of native FP64 performance. CUDA 13.5 further optimizes matrix-multiplication heuristics and kernel tuning for Vera Rubin’s NVFP4 and MXFP8 operations, with NVIDIA claiming improved throughput without requiring per-shape autotuning.

Enhanced GPU Scheduling and Resource Management

CUDA 13.5 introduces new features for GPU workload management, including:

• Locality domain enhancements: Applications can now choose between uniform streaming multiprocessor (SM) partitions or partitions that include all available SMs, maximizing GPU hardware utilization.
• NVIDIA Multi-Process Service (MPS) v3: Weighted timeslicing allows workloads to receive longer compute turns, while SM partition load balancing enables borrowing and lending of resources between partitions.
• Cgroup-based routing: Workload launchers can map process groupings to MPS resource policies, streamlining resource allocation.

Expanded Platform and Developer Tool Support

CUDA 13.5 extends Windows on Arm support to NVIDIA DGX Station for Windows, building on its GA status in CUDA 13.4. The release also introduces public preview support for CUDA Tile to Single Instruction Multi-Thread (SIMT) interoperability, enabling CUDA Tile kernels to call SIMT library functions directly. Additionally, multi-GPU communication is now supported within CUDA Tile kernels, facilitating advanced workloads such as fused matrix multiplication and cross-GPU reduction.

Deployment and Compatibility Updates

CUDA 13.5 adjusts deployment compatibility for Rubin and later GPUs, which now support Windows Display Driver Model (WDDM) and Microsoft Compute Driver Model (MCDM) but no longer support Tesla Compute Cluster (TCC) mode. Developers must adapt applications and deployment scripts to these supported configurations. The release also includes updates to the Nsight Perf SDK for AI, enabling metric collection and workload analysis in Perfetto.

As NVIDIA continues to refine its CUDA platform, the release of CUDA 13.5 underscores its commitment to advancing GPU computing capabilities for scientific and enterprise applications.