NVIDIA CUDA 13.5 is here, bringing a substantial set of updates for developers working with artificial intelligence, GPU accelerated applications and high performance computing. The latest release introduces general availability support for NVIDIA’s upcoming Vera Rubin architecture, alongside improvements to GPU scheduling, CUDA programming tools and developer libraries.
The update also expands support for Windows on Arm, enhances NVIDIA Multi Process Service (MPS), and introduces new capabilities for CUDA Tile and CUDA Python. Together, these changes aim to make GPU resources easier to manage while giving developers greater flexibility when building demanding AI workloads.
One of the biggest highlights is the arrival of Vera Rubin support at the general availability stage. Developers can now begin preparing production workloads for NVIDIA’s next generation computing platform, rather than relying solely on early access development support.
CUDA 13.5 also focuses on improving the way applications use available GPU resources. From more flexible scheduling to advanced multi GPU communication, the release offers several features designed to make complex computing workloads more efficient.
NVIDIA Vera Rubin support reaches general availability
The most notable addition in CUDA 13.5 is general availability support for NVIDIA Vera Rubin. The transition from early developer access to GA support represents an important milestone for developers preparing applications for NVIDIA’s next generation hardware.
CUDA 13.5 also extends Windows on Arm support, which first reached general availability in CUDA 13.4, to NVIDIA DGX Station for Windows.
Another important change concerns Windows driver support on Rubin based GPUs. CUDA 13.5 supports both the Windows Driver Display Model (WDDM) and Microsoft Compute Driver Model (MCDM).
MCDM provides a compute focused driver model for supported workloads. However, NVIDIA has confirmed that Tesla Compute Cluster (TCC) mode will not be supported on Rubin and subsequent GPU architectures.
This means developers and deployment teams relying on TCC will need to review their configurations before moving to newer NVIDIA hardware.
The update is particularly relevant for enterprise users and AI infrastructure teams planning their next GPU deployments. Having official CUDA support available should make it easier to prepare software environments and production applications for the new architecture.
CUDA 13.5 improves GPU scheduling and resource management
NVIDIA is giving developers more control over how GPU resources are allocated and shared between applications.
One of the major improvements involves locality domains, which allow applications to organise GPU resources according to their hardware locality.
CUDA 13.5 introduces additional flexibility when dividing streaming multiprocessors (SMs) across these domains. Developers can now choose between uniform partitions, which provide predictable resource allocation, and partitions that use all available SMs within each locality domain.
The new supplemental SM resource option allows applications to include additional GPU resources when creating partitions. This can improve hardware utilisation, although the resulting partitions may not all have identical sizes.
Applications can also use the updated resource discovery functionality to determine how many SMs are available in each partition and distribute workloads accordingly.
This gives developers a choice between consistent partition sizes and maximising the GPU resources available to their applications.
NVIDIA Multi Process Service also receives several important improvements.
CUDA 13.5 introduces more flexible weighted timeslicing, allowing developers to control how much GPU execution time different workloads receive.
This could be particularly useful when several AI workloads share the same GPU and have different performance requirements.
The update also introduces SM partition load balancing. Workloads can temporarily borrow additional GPU resources when capacity is available, while unused resources can be made available to other partitions.
NVIDIA has added cgroup based routing and namespace level SM partition assignments as well. These changes should simplify GPU resource management in environments where multiple applications, containers and users share computing infrastructure.
CUDA Tile and CUDA Python receive important upgrades
CUDA Tile receives several improvements aimed at making GPU programming more flexible, particularly for multi GPU AI workloads.
CUDA 13.5 introduces support for communication between GPUs directly within CUDA Tile kernels.
Developers can use unicast pointers to access memory on another NVLink connected GPU. Multicast pointers and dedicated multimem operations also enable communication patterns such as reductions and stores across multiple GPUs connected through supported NVIDIA NVSwitch infrastructure.
This could be useful for workloads that combine computation and communication, including advanced matrix multiplication and cross GPU reduction operations.
Another important addition is Tile to SIMT interoperability, which arrives as a public preview.
Previously, CUDA Tile and SIMT library code generally had to be combined across separate kernel boundaries. The new preview allows CUDA Tile kernels to call supported SIMT library functions directly.
This means developers can potentially avoid unnecessary data transfers between kernels and improve the flexibility of their GPU applications.
The preview supports selected libraries, including NVIDIA NVSHMEM, CCCL Cooperative, NVIDIA Warp and NVIDIA MathDx.
CUDA Python is also becoming more capable.
The CUDA Python core library, cuda.core 1.2, introduces batched memory copies, Programmatic Dependent Launch support, simplified runtime compilation and improved CUDA Graph updates.
Batched memory copies allow developers to submit multiple transfers through a single CUDA driver call. This could reduce overhead in applications that frequently move data between memory buffers.
Programmatic Dependent Launch provides additional control over how dependent GPU kernels execute, allowing certain workloads to overlap operations more efficiently on supported hardware.
CUDA Graph updates also become more flexible. Developers can modify supported kernel parameters, memory operations and other graph elements without necessarily rebuilding the entire executable graph.
These improvements could make CUDA Python more attractive for AI researchers and developers who want to access GPU capabilities without relying exclusively on traditional C++ workflows.
CCCL and NVIDIA developer tools get smarter
CUDA Toolkit 13.5 also brings CCCL 3.5, with several improvements aimed at AI workloads and performance tuning.
Developers can now access public tuning APIs for CUB device wide algorithms. These allow more control over parameters such as threads per block, items per thread, vectorisation and algorithm selection.
The release also introduces batched Top K selection, which is particularly relevant to AI inference and token selection during text generation.
The new operation supports selecting the largest or smallest K values across multiple segments, including scenarios where segment sizes and K values vary.
Another improvement introduces run to run reproducibility for floating point scans. Developers can request consistent results for the same GPU, input and build configuration.
CUDA 13.5 also allows certain reductions to use element counts stored directly in GPU accessible memory. This can reduce the need to transfer information back to the CPU, helping applications keep data dependent operations on the GPU.
NVIDIA is also expanding its developer profiling tools.
The NVIDIA Nsight Perf SDK for AI enters public preview, offering profiling capabilities for long running AI training and inference workloads.
The SDK preserves selected CPU and GPU performance information from the original execution, allowing developers to investigate performance problems that may be difficult to reproduce.
Meanwhile, Nsight Compute 2026.4 introduces a dedicated timeline page that helps developers examine GPU activity over time and identify performance bottlenecks.
Overall, CUDA 13.5 is a substantial update for NVIDIA’s developer ecosystem. Vera Rubin support is the headline addition, but the improvements to scheduling, multi GPU communication, Python development and profiling tools could prove just as important for everyday development.
For AI developers and enterprise computing teams, the release offers more control over GPU resources and additional tools for building applications that can take advantage of NVIDIA’s evolving hardware platform.
Follow TechBSB For More Updates
