NVIDIA CUDA Toolkit 13.5: Vera Rubin Support, Windows Driver Changes and Multi-GPU Tools

What shipped​

NVIDIA's technical blog says CUDA Toolkit 13.5 reaches general availability with production-ready Vera Rubin support. It also extends Windows on Arm support to NVIDIA DGX Station for Windows. The CUDA Toolkit Archive page lists CUDA Toolkit 13.5.0 (November 2026) as the latest release. NVIDIA's blog says Windows on Arm support has been generally available since CUDA 13.4. What's new is the extension to DGX Station for Windows.

VideoCardz framed the release as a multi-GPU and AI-workload update. NVIDIA's own documentation backs that up. The "What's New" list highlights Vera Rubin GA, multi-GPU communication from CUDA Tile kernels, and CUDA Python additions such as Programmatic Dependent Launch, batched memcopies and more flexible CUDA Graph updates.

The Windows angle: WDDM, MCDM and the end of TCC on Rubin​

NVIDIA's blog says that, starting with Rubin, CUDA supports the Windows Driver Display Model (WDDM) and Microsoft's Compute Driver Model (MCDM). MCDM is the compute-only option. Tesla Compute Cluster (TCC) mode is not supported on Rubin and later GPUs. If your deployment scripts or tooling assume TCC for compute cards, NVIDIA says to move to a supported WDDM or MCDM configuration on Rubin-class hardware.

Drivers: running versus using the new features​

NVIDIA's release notes make a distinction that is easy to miss.

• Existing CUDA 13.x applications: they keep running on drivers 580 or later through minor-version compatibility.
• New 13.5 features and newly enabled platforms: these need an R620-branch or later driver that supports them.
• RTX Spark on Windows: the release notes give driver 616.41 or later.
• Bundling: the driver is no longer included with the toolkit. NVIDIA says this applies on Windows starting with CUDA 13.1 and on Linux starting with 13.4, so you install the driver separately.

Admins should treat the toolkit and driver as two separate rollouts. Updating the toolkit on a build machine does not mean every deployment target can use the new features. The companion Linux driver for this release is reported as 620.70.06, according to pbxscience.

Sharing a GPU: MPS v3 and locality domains​

Most of the infrastructure-flavored changes are in Multi-Process Service (MPS) v3. NVIDIA's blog describes these additions:

• Weighted timeslicing. You set a base timeslice, scale it per MPS server, and tune how aggressively the GPU switches between workloads. Servers hosting demanding workloads get longer turns.
• SM partition load balancing. A workload can borrow streaming multiprocessors beyond its base partition when they're free. Borrowing is enabled with --max-chunks. A partition with no attached clients can lend its SMs with --shareable=true. The two settings are independent. MPS reclaims borrowed resources before reassigning them, and applications still see a stable base partition size.
• Cgroup-based routing. You map a cgroup v2 subtree to an MPS namespace, and clients connecting through the daemon's main socket are routed there automatically.
• Per-namespace SM partition assignments. You pick a partition per device once, so you no longer set CUDA_MPS_SM_PARTITION for each client. Namespace assignments win, and devices without one fall back to the environment variable.
• Polling-based memory accounting. If cgroup threshold notifications are missing or unreliable, you can run nvidia-cuda-mps-control memacct set force_polling=true before accounting starts. NVIDIA notes this adds work per eviction decision.

Borrowing is opportunistic. It improves utilization when capacity is idle and is not a guarantee of extra speed.

Locality domains offer a related trade-off. By default, localized green-context partitions get the same guaranteed SM count per domain. A new flag, cudaDevSmResourceGroupLocalityDomainSupplementalSms, includes the leftover SMs. That uses more of the GPU but can produce unequal partitions. Developers can use discovery mode in cudaDevSmResourceSplit() to read each partition's SM count and balance work accordingly.

CUDA Tile, Python and CCCL​

• CUDA Tile: kernels can now communicate across GPUs. Unicast pointers reach memory on NVLink-connected GPUs. Multicast pointers and multimem operations cover reductions and stores on NVSwitch fabrics with NVLink SHARP.
• Tile and SIMT interoperability (public preview): Tile kernels can call SIMT library device functions directly. NVIDIA names NVSHMEM, CCCL Coop, Warp and MathDx, and says each library will enable this support soon. Calling your own __device__ functions from tile code is planned for a future release.
• CUDA Python (cuda.core 1.2): copy_batch() submits several buffer copies in one driver call. Source and destination sizes must match, and copies must not alias. PDL is exposed through LaunchConfig on compute capability 9.0 or newer. Graph nodes can be updated without reinstantiating the graph.
• CCCL 3.5: it adds public CUB tuning APIs, reproducible floating-point scans on the same GPU with the same input and build configuration, and batched Top-K selection, which is integrated into FlashInfer. Reductions can read an element count from device memory, so a captured graph can handle a different count on each replay.

Libraries, tools and packaging​

NVIDIA reports speedups in cuSPARSE, all of them NVIDIA's own measurements:

• CSR SpMM (ALG3) is 1.65x to 2.27x faster across data-center GPUs.
• SpSV ALG2 is 2.7x faster than ALG1.
• SpSV analysis is 2.79x faster.
• csrcolor is roughly 3x faster.

The release notes also report up to 212 TFLOPS of emulated DGEMM on Rubin through Ozaki-based emulation. Your own workloads will vary.

On the tools side:

• Nsight Perf SDK for AI: a public preview of always-on profiling for long-running training and inference, with data explorable in Perfetto.
• Nsight Compute 2026.4: adds a timeline page.
• CUDA-GDB: can stop on filtered CUDA Driver log events.
• nvprune --precompile: compiles PTX or TileIR to native SASS/ELF at prune time, which avoids runtime JIT overhead for known targets. NVIDIA advises validating the output on every architecture you support.
• cuobjdump: symbol-table dumps now show descriptive attributes in additional columns. If you parse that output with scripts, update the parsers.

Known issues and caveats​

The release notes list some items worth reading before you roll out:

• Support for the nvidia-peermem.ko kernel module is deprecated and will be removed in the upcoming CUDA Toolkit release.
• A cuBLASLt issue with NVFP4 on compute capability 10.x GPUs affects only manually overridden algorithm configurations. It was introduced in 13.5.
• A rare Ozaki-II double-complex BLAS issue has been present since 13.4. It involves nearly collinear vectors.

Should you upgrade?​

Upgrade now if you need Vera Rubin, DGX Station for Windows or the new MPS controls. Otherwise, test first. Check that your driver meets the feature-specific requirement, rebuild against the new toolkit in CI, and look for anything that parses tool output. The Tile-SIMT interop and the Nsight Perf SDK for AI are previews, so treat them as such in production planning.

If you're a Windows gamer with an existing GeForce card, this release changes nothing you'll notice. Existing CUDA 13.x applications keep working on drivers 580 and later.