NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support
NVIDIA CUDA 13.5 delivers general availability for the NVIDIA Vera Rubin GPU architecture, enabling enterprise production use. The release also extends Windows on Arm support to the NVIDIA DGX Station for Windows, offering new scheduling and memory management features.
Tags
Quick summary
NVIDIA CUDA 13.5 delivers general availability for the NVIDIA Vera Rubin GPU architecture, enabling enterprise production use. The release also extends Windows on Arm support to the NVIDIA DGX Station for Windows, offering new scheduling and memory management features.
NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support
As artificial intelligence and accelerated computing continue to push the boundaries of what is possible, the latest NVIDIA CUDA release arrives to unlock unprecedented performance and scale. NVIDIA CUDA 13.5 delivers general availability (GA) across major new platforms, marking a significant milestone for enterprise production environments.
According to NVIDIA, this release moves support for the NVIDIA Vera Rubin GPU architecture from early developer access to general availability. Additionally, Windows on Arm support, which became GA since CUDA 13.4, now extends specifically to the NVIDIA DGX Station for Windows. These platform expansions ensure that developers can deploy workloads on a wider array of hardware configurations with production-ready stability.
Locality Domain Enhancements and Resource Partitioning
A core focus of CUDA 13.5 is the increased customization of how GPU work is scheduled via locality domains. Applications now possess a choice between uniform streaming multiprocessor (SM) partitions across locality domains or partitions that include all available SMs.
By default, localized green-context partitions utilize the same guaranteed SM count per domain, leaving supplemental SMs outside those partitions. However, the new cudaDevSmResourceGroupLocalityDomainSupplementalSms flag, combined with cudaDevSmResourceGroupLocalityDomainId, allows applications to include these additional SMs within each selected domain. This enables the utilization of all device SMs when partitioning across locality domains, although the resulting partitions may differ in size.
To facilitate this flexibility, applications can use discovery mode in the cudaDevSmResourceSplit() API. This API allows developers to obtain each partition's SM count and distribute work accordingly. As noted in the documentation, this capability allows users to choose between the uniformity and predictability of uniform SM size for each locality domain, or to choose the maximum number of SMs that each locality domain has to utilize as much of the GPU hardware as possible.
NVIDIA Multi-Process Service v3 Innovations
NVIDIA Multi-Process Service (MPS) v3 introduces significant control over how workloads share a GPU, ranging from scheduling individual servers to assigning resources to groups of clients. Several new features enhance resource management and efficiency.
Weighted Timeslicing and Load Balancing
Weighted timeslicing allows administrators to give longer compute turns to servers hosting workloads that require them. Users can set a base timeslice, scale it per MPS server, and independently tune how aggressively the GPU switches between workloads.
SM partition load balancing enables workloads to use capacity beyond their base partition when resources are available. This feature includes borrowing and lending mechanisms. Users can enable borrowing with --max-chunks and allow a partition to lend its SMs while it has no attached clients with --shareable=true. These settings are independent, meaning a partition can borrow, lend, or do both. MPS reclaims borrowed resources before reassigning them, while applications continue to see a stable base partition size.
Cgroup-Based Routing and Namespace Assignments
Cgroup-based routing connects a workload launcher's process groupings to MPS resource policies. By mapping a cgroup v2 subtree to an MPS namespace, clients connecting through the daemon's main socket are routed to that namespace automatically.
Furthermore, per-namespace SM partition assignments allow administrators to choose a partition for each device once and apply that selection to every client routed to the namespace. This eliminates the need to configure CUDA_MPS_SM_PARTITION separately for those clients. Namespace assignments take precedence, while devices without an assignment retain the environment-variable fallback.
Polling-Based Memory Accounting
To improve reliability in memory management, a polling-based fallback for memory accounting is introduced. This feature provides a fallback when cgroup threshold notifications are missing or unreliable. During eviction decisions, MPS can read current usage and limits directly instead of relying on cached notification state. To force polling, users can run nvidia-cuda-mps-control memacct set force_polling=true before memory accounting starts. NVIDIA notes that this adds work per eviction decision but removes dependence on notification delivery.
CUDA Tile and Multi-GPU Communication
CUDA 13.5 adds support for multi-GPU communication directly within CUDA Tile kernels. Unicast pointers allow a kernel to access memory on another NVLink-connected GPU. Additionally, multicast pointers and dedicated multimem operations support reductions and stores across multiple GPUs on NVIDIA NVSwitch fabrics with NVIDIA NVLink SHARP support.
Applications prepare the shared-memory mappings before launch, then combine communication with computation inside the kernel. This capability enables workloads such as fused matrix multiplication and cross-GPU reduction.
SIMT Interoperability Preview
In a public preview, CUDA 13.5 allows CUDA Tile kernels to call SIMT library device functions directly. Previously, Tile and SIMT library code could only be combined across kernel boundaries, with data passing through global memory between them. With this update, kernels written in cuTile Python or CUDA Tile C++ can call device functions from SIMT libraries in the middle of their own code, on data that is already in registers or shared memory.
Supported libraries include NVIDIA NVSHMEM, NVIDIA CCCL Coop, NVIDIA Warp, and NVIDIA MathDx. The latter covers NVIDIA cuFFTDx, NVIDIA cuBLASDx, NVIDIA cuSolverDx, and NVIDIA nvCOMPDx. NVIDIA states that this interop support will be enabled by each of these libraries soon.
To enable the preview in C++, users compile with nvcc --tile-preview simt and pass the tilelib to the tile compiler. In cuTile Python, the cuda.tile_preview.foreign_call function, provided by the cuda.tile_preview package, is used. NVIDIA clarifies that in version 13.5, the Tile-SIMT interop preview covers library functions only; calling custom __device__ functions from tile code is planned for a future release.
Windows Driver Support on Rubin
Starting with the NVIDIA Rubin architecture, CUDA supports the Windows Driver Display Model (WDDM) and the Microsoft Compute Driver Model (MCDM). MCDM provides the compute-only driver model. NVIDIA explicitly states that Tesla Compute Cluster (TCC) is not supported on Rubin and later GPUs. Consequently, applications and deployment tools that assume TCC mode should use the supported WDDM or MCDM configuration for their target system.
CUDA Python and CCCL Updates
NVIDIA CUDA Python continues to expand Pythonic access to CUDA programming. The CUDA Python core library cuda.core 1.2 introduces batched memory copies, Programmatic Dependent Launch (PDL), simpler runtime compilation, and more flexible CUDA Graph updates.
The new cuda.core.utils.copy_batch() submits multiple buffer copies through a single CUDA driver call. CopyOptions provides control over source access ordering, memory-location hints, and overlap behavior. For example, the following code submits two independent copies using existing source and destination buffers:
from cuda.core.utils import copy_batch
# Each source and destination pair must have matching sizes.
# Copies within the batch must not alias.
copy_batch(stream, srcs=[src_a, src_b], dsts=[dst_a, dst_b])
stream.sync()LaunchConfig now exposes PDL through the programmatic_stream_serialization option. On GPUs with compute capability 9.0 or newer, cooperating kernels can overlap execution within the same CUDA stream, allowing a dependent kernel to begin independent work before its predecessor completes.
Applications can also update graph-definition nodes before future instantiations and modify executable graph nodes through graph[node] views without reinstantiating the graph. Supported updates include kernel parameters, memory copies, memory sets, child graphs, events, and host callbacks.
In the CCCL 3.5 release, public tuning APIs are provided for CUB. The update also includes reproducible floating-point scans and batched Top-K selection integrated into FlashInfer. Additionally, data-dependent reductions consume element counts from device memory using cuda::args::deferred.
Conclusion
NVIDIA CUDA 13.5 represents a substantial evolution in the toolkit for accelerated computing, particularly with the general availability of the NVIDIA Vera Rubin architecture. By introducing granular control over locality domains, enhancing MPS v3 with weighted timeslicing and cgroup routing, and enabling direct multi-GPU communication within CUDA Tile kernels, the release empowers developers to optimize workloads for scale and efficiency.
The integration of Windows on Arm support for the DGX Station and the shift to WDDM/MCDM models for Rubin further broadens the deployment landscape. As NVIDIA continues to refine these capabilities through public previews and library updates, the ecosystem remains poised to support the next generation of AI infrastructure.



