The news

Primary: Jonathan Bentz, CUDA Toolkit 13.4 Adds Windows on Arm Support and Greater Control over Shared GPUs, NVIDIA Developer Blog, 9 September 2026.

NVIDIA released CUDA Toolkit 13.4. The operator-facing headlines are platform reach, Rubin preview, shared-GPU control, and a default change on coherent CPU-GPU machines.

What changed

Windows on Arm. CUDA on Arm was already a Linux story. 13.4 adds Windows on Arm so the same CUDA apps can target that client/server path.

Rubin preview. Functional support for NVIDIA Rubin at compute capability 107, with NVCC SM_107. NVIDIA frames this as early porting ahead of general availability in a future toolkit.

MPS V3. Multi-Process Service gets a modernized control layer: scriptable CLI, named server instances, namespaces, TOML config, SM partition controls, and cgroup-integrated GPU memory limits. Aimed at precise partitioning inside containers.

Fabric and locality. CUDA Compute Fabric Transport (CFT) adds endpoint-ID based put/get/reduce across NVLink fabric for communication-library authors. Locality domains let apps co-locate memory allocations and green-context SMs. Unified-memory residency queries (cudaMemGetLocationInfo) expose where managed data currently lives.

Install and memory defaults.

  • SDK installers no longer bundle the NVIDIA driver. Install driver and toolkit as separate packages.
  • On coherent platforms (Grace Hopper, Grace Blackwell, Vera Rubin), the driver defaults to CDMM instead of NUMA. NUMA remains supported via a kernel-module parameter. The setting is node-wide and requires a driver reload or reboot. NVIDIA says choose the mode before upgrading.

Toolchain and libraries. Host compatibility adds GCC 16 and Clang 22. CUDA Python (cuda.core 1.1.0, cuda.compute 1.1) and CCCL 3.4 ship with the toolkit. Nsight tools and core math libraries gain CUDA 13.4 / Rubin / Windows-on-Arm coverage as described in the primary.

Who must act

  • Cluster operators on Grace / Vera Rubin coherent nodes: decide CDMM vs NUMA before the driver bump; document the module parameter and reboot window.
  • Platform teams with shared GPUs: evaluate MPS V3 TOML + cgroup limits vs current MPS or MIG practice.
  • ISVs preparing Rubin: start preview builds with sm_107; keep GA expectations tied to a future toolkit.
  • Windows on Arm ISVs: add 13.4 to the support matrix.
  • CI/CD owners: stop assuming the toolkit installer drops a driver.

What's new vs prior CUDA 13.x

  1. WoA support.
  2. Rubin sm_107 preview.
  3. MPS V3 control plane.
  4. CFT + locality domains + residency queries.
  5. Driver decoupled from SDK installer.
  6. CDMM as default on named coherent platforms.

What it does not settle

  • Rubin GA toolkit version and date: UNKNOWN.
  • Whether your app silently depends on NUMA behavior on coherent nodes: measure before fleet upgrade.
  • Production isolation SLOs for MPS V3 under Kubernetes: UNKNOWN beyond NVIDIA docs.
  • Relationship to CUDA Rust tracks: separate release narrative (see live Stack cuda-rust-two-tracks).

What to do now

  1. Read the CDMM section of the primary and the linked CDMM explainer before upgrading coherent nodes.
  2. Inventory automation that installs toolkit+driver as one blob; split the packages.
  3. Pilot MPS V3 on a non-prod shared GPU pool with explicit memory limits.
  4. If you need Rubin ports, enable SM_107 in a preview branch only.
  5. For deeper pick-vs / failure modes, see Stack draft cuda-13-4-operator-stack and Field Notes cuda-13-4-upgrade-watchlist.