NVIDIA CUDA 13.4 Toolkit Arrives With Windows-on-Arm Support and Early Vera Rubin Access
NVIDIA has rolled out the CUDA 13.4 Toolkit, a major update for developers working on GPU-accelerated applications, AI workloads, and next-generation computing platforms. The latest release expands CUDA support to Windows-on-Arm systems, prepares the software ecosystem for upcoming RTX Spark PCs, and introduces preview support for NVIDIA’s future Vera Rubin CPU and GPU platform.
One of the biggest additions in CUDA 13.4 is support for Windows on Arm64. This is an important step as NVIDIA prepares for the arrival of RTX Spark Windows laptops, expected in October. With this update, developers can begin porting, testing, and optimizing their applications for Windows Arm64 before the new RTX Spark systems reach users.
The new support also allows CUDA to run on existing Windows-on-Arm machines, giving developers an early environment to validate software paths, check dependencies, and prepare CUDA-based applications for a broader range of devices. For teams building AI tools, creative applications, simulation software, or GPU-accelerated workflows, this gives them more time to ensure compatibility and performance from day one.
CUDA 13.4 is also closely tied to NVIDIA’s upcoming RTX Spark roadmap. By making the toolkit available ahead of the hardware launch, NVIDIA is giving software developers the tools needed to get their applications ready before the new platform becomes widely available. This can help reduce compatibility issues and make the transition smoother for both developers and end users.
Another major highlight of CUDA 13.4 is early developer access for NVIDIA’s Vera Rubin CPUs and GPUs. The toolkit adds preview functional support for the Rubin GPU architecture, listed as compute capability 107. This allows developers and library maintainers to begin preparing their software before full Rubin support becomes generally available in a future CUDA release.
Rubin is expected to play a key role in NVIDIA’s next wave of AI computing, particularly in the growing field of agentic AI. By offering early access through CUDA 13.4, NVIDIA is giving developers a head start on porting applications, testing code, and validating performance on the next-generation architecture.
CUDA 13.4 also introduces a wide range of platform improvements across the CUDA compiler, PTX ISA, CUDA C++, CUDA Tile programming, runtime APIs, libraries, and developer tools. These updates are designed to improve flexibility, performance, and developer productivity across different GPU computing environments.
The release includes Multi-Process Service V3, which brings a modernized control layer for managing GPU workloads. It adds a scriptable command-line interface, named server instances, TOML configuration support, and cgroup-integrated GPU memory limits. These features are especially useful for containerized environments where developers need more precise GPU partitioning and resource management.
CUDA Compute Fabric Transport is another important addition. It provides a transport-focused API designed for communication-library developers. This makes it easier to move data across NVLink fabric using named logical endpoints and asynchronous operations, helping improve performance in advanced multi-GPU and high-performance computing setups.
CUDA Python also receives notable upgrades in this release. The cuda.core 1.1.0 update adds support for texture and surface programming, NUMA-aware managed memory, and .pyi type stubs. Meanwhile, cuda.compute 1.1 introduces ahead-of-time compilation of CCCL algorithms for multiple GPU architectures. These improvements should make CUDA Python more capable and convenient for developers working in AI, scientific computing, and data-heavy applications.
NVIDIA has also updated core math libraries in CUDA Toolkit 13.4 with functional support for the Rubin GPU architecture. In addition, these libraries now include Windows-on-Arm support for the N1X laptop ecosystem, further strengthening CUDA’s reach across emerging Arm-based Windows devices.
Overall, CUDA 13.4 is a forward-looking release that prepares NVIDIA’s developer ecosystem for several major platform shifts. Windows-on-Arm support opens the door for CUDA acceleration on a new class of Windows devices, RTX Spark readiness gives developers time to optimize before launch, and Vera Rubin preview support lays the groundwork for NVIDIA’s next generation of AI hardware.
For developers building GPU-accelerated applications, CUDA 13.4 offers an important early opportunity to test, validate, and optimize software for the platforms that will shape NVIDIA’s computing roadmap in the months ahead.
