NVIDIA CUDA Toolkit 13.4 Adds Windows on Arm, Rubin GPU Support

NVIDIA has rolled out CUDA Toolkit 13.4, marking a significant upgrade for developers leveraging GPU acceleration. The update introduces support for Windows on Arm platforms, early developer access to the Rubin GPU architecture, and enhanced GPU management tools, broadening the appeal of CUDA for high-performance computing and AI workloads.

CUDA’s expansion to Windows on Arm is a key milestone, addressing growing demand for Arm-based devices in edge computing and mobile environments. Historically, CUDA supported Arm through Linux only, but this release extends compatibility, opening new opportunities for developers targeting lightweight and energy-efficient systems.

Another headline feature is the preview support for NVIDIA’s Rubin architecture, which offers compute capability 107. Rubin is poised to power the next generation of AI workloads, focusing on agentic AI systems. Developers can now begin porting applications to Rubin GPUs ahead of its general availability, making this release especially relevant for teams preparing to adopt cutting-edge hardware.

Advanced GPU Resource Management

The toolkit also introduces Multi-Process Service (MPS) V3, a revamped system for managing shared GPU resources. This upgrade includes a scriptable command-line interface, named server instances, and GPU memory limits integrated with Linux cgroups. The enhancements enable precise partitioning of GPU resources, critical for applications requiring strict isolation in containerized environments, such as enterprise AI and HPC workloads.

CUDA Compute Fabric Transport (CFT), another new feature, offers advanced mechanisms for moving data across GPUs at scale. By reducing virtual address space pressure and supporting asynchronous operations, CFT improves data transfer efficiency in large multi-GPU systems, making it a vital tool for developers building on NVIDIA’s NVLink technology.

Python and Developer Tools

Python developers gain expanded access to CUDA APIs through updates to CUDA Python, including improved memory management and support for Pythonic texture programming. CUDA core libraries now enable advanced workflows, such as NUMA-aware memory management and ahead-of-time compilation for GPU algorithms. These updates simplify the process of developing high-performance applications while maintaining flexibility across multiple GPU architectures.

NVIDIA’s developer tools have also been enhanced. Nsight Python 1.0 introduces streamlined kernel profiling for Python workflows, while Nsight Systems now supports Rubin GPUs and CUDA 13.4 across expanded platforms. These tools provide critical debugging and optimization capabilities for developers targeting complex workloads in AI, HPC, and cloud environments.

Implications for AI and HPC Developers

CUDA Toolkit 13.4’s focus on GPU resource management and advanced communication protocols underscores its growing importance in high-performance AI and HPC deployments. With companies increasingly adopting multi-GPU setups for tasks like deep learning and data analysis, the toolkit’s updates will likely accelerate adoption in enterprise and research contexts.

Moreover, Rubin GPU preview support aligns with NVIDIA’s broader push into AI-driven solutions, following its recent expansions with the NVIDIA Agent Toolkit and CUDA-X libraries. These developments position NVIDIA to capitalize on the surge in demand for AI infrastructure, which is projected to grow significantly in the coming years.

Market Context

NVIDIA’s stock (as of September 9, 2026) is trading at $223.67, down 0.87% over the past 24 hours, with a market cap of $5.43 trillion. While the immediate impact of this release on NVIDIA’s financials may be muted, the long-term implications for its AI and HPC market share could be substantial, especially as Rubin GPUs transition to full availability.

Developers and enterprises leveraging CUDA Toolkit 13.4 are poised to gain a competitive edge, especially in AI and machine learning applications requiring scalable, high-performance compute power. For those planning to adopt Rubin GPUs or scale workloads on Arm devices, this release provides essential tools to start optimizing their applications today.

To explore the full list of features, developers can access the toolkit and release notes on NVIDIA’s official site.

Image source: Shutterstock