CUDA Toolkit 13.4 is out, and NVIDIA has packed this one with changes that reach from developer laptops to multi-GPU datacentres. The headline is straightforward: CUDA applications can now target Windows on Arm, while developers can begin porting work to NVIDIA’s Rubin architecture before full toolkit support arrives.
NVIDIA Watch analysis
What happened
NVIDIA announced the release on 9 September 2026. Alongside the platform expansion, it introduces Multi-Process Service V3, a more programmable way to manage shared GPU resources in containerised environments, plus new transport primitives for software moving data across NVLink fabric.
Our top picks
- Windows on Arm support
CUDA development is no longer limited to Linux on Arm when the target platform is Windows. - Rubin preview support
Developers can begin porting applications for Rubin, using the new compute capability 107 target ahead of general availability. - MPS V3 resource controls
A scriptable CLI, named servers, TOML configuration and cgroup memory limits make shared GPUs easier to partition and automate. - CUDA Compute Fabric Transport
Communication-library developers can use named logical endpoints and asynchronous operations to move data across NVLink. - A more capable CUDA Python
The 1.1.0 release adds texture and surface programming, richer managed-memory controls, graph improvements and type stubs for public APIs. - Ahead-of-time GPU algorithm builds
The 1.1 compute library can compile CCCL algorithms for several GPU architectures without needing a GPU on the build machine. - A faster Blackwell scan
NVIDIA says the new CCCL 3.4 DeviceScan implementation reaches up to 92% memory-bandwidth utilisation in tested workloads.
Why it matters
This is less a shiny single feature than a widening of CUDA’s useful territory. Windows on Arm lowers a platform barrier, Rubin preview support gives developers lead time, and MPS V3 tackles the unglamorous but vital problem of making expensive shared accelerators behave neatly in containers.
The release also gives Python developers more of the CUDA stack, while library authors get lower-level tools for portability and multi-GPU communication. The 92% figure is NVIDIA’s benchmark result for specified Blackwell scan workloads, not a general promise about every CUDA programme.
Our read
CUDA 13.4 looks like infrastructure work with unusually visible consequences. Developers targeting Windows on Arm or preparing for Rubin should investigate now; most application teams can sensibly wait for the higher-level libraries and hardware support they already use to absorb the changes.
What to watch
- Whether Windows on Arm support attracts a meaningful developer ecosystem.
- When Rubin support moves from preview to general availability.
- How much operational friction MPS V3 removes in real container deployments.
- Whether the new Python and CCCL paths become production defaults or remain specialist tools.
Discussion spark: Which CUDA 13.4 change is most likely to affect your work first: Windows on Arm, Rubin preview support, shared-GPU controls or the expanded Python tooling?
Sources and evidence
- CUDA Toolkit 13.4 Adds Windows on Arm Support and Greater Control over Shared GPUs (9 September 2026, 20:24 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.