NVIDIA released CUDA 13.5 with general availability support for Vera Rubin, moving beyond the earlier developer stage. The update also changes how graphics processors share workloads and exchange data, while expanding Python and debugging tools. Some additions, including Tile access to existing library functions and a new AI profiling kit, remain in public preview rather than sharing the platform support’s production status.
NVIDIA released CUDA 13.5 on November 4, bringing production-ready support for its Vera Rubin graphics-processor platform. CUDA is the company’s software toolkit for programming accelerated computing workloads. Earlier CUDA 13 versions offered developer-stage support for Vera Rubin. The new release also extends Windows on Arm support to DGX Station, NVIDIA’s workstation platform for AI development.[1], [2]
NVIDIA Multi-Process Service, or MPS, manages the sharing of a graphics processor between processes. Its version 3 adds weighted time allocation, letting administrators give some workloads longer computation turns while separately adjusting the frequency of switching between tasks. A partition can also lend idle processing resources or borrow capacity beyond its own base allocation when resources are available.[1]
Borrowed capacity is reclaimed before being reassigned, while applications continue to see a stable base partition size. Process groups can be directed to MPS resource policies through cgroup v2. Locality-domain controls offer another choice: applications can use uniformly sized partitions or include additional processing units, with discovery interfaces reporting how many units each partition actually receives.[1]
CUDA Tile, NVIDIA’s programming approach for blocks of data, adds communication across graphics processors linked by NVLink and NVSwitch. Operations can address one remote processor or multiple recipients. A separate public preview lets Tile code call functions from libraries built for single-instruction, multiple-thread execution, known as SIMT. That preview covers library functions; arbitrary user-defined functions are outside its present scope.[1]
Python users receive batched memory copying, dependent launches and graph updates through cuda.core 1.2. The CUDA C++ Core Libraries package, CCCL 3.5, expands tuning and repeatable floating-point scans under the same processor, input and build conditions. Nsight Compute 2026.4 adds a timeline for examining workload execution. NVIDIA’s Nsight Perf software development kit for AI enters public preview for investigating long-running training and inference activity.[1]
