NVIDIA Releases CUDA Toolkit 13.4 with Windows on Arm Support and Rubin Preview

CUDA Support Arrives for Windows on Arm

CUDA applications have historically relied on Linux for Arm platform deployment, but the new release expands that capability directly to the Windows on Arm ecosystem, according to NVIDIA documentation. Writing for Wccftech, Senior Editor Hassan Mujtaba notes that this update enables developers to prepare Windows Arm64 applications ahead of the scheduled October rollout of RTX Spark laptops. Developers can now validate dependencies and test software paths on existing Windows on Arm machines, according to NVIDIA. The expansion aligns with the upcoming hardware releases for the N1X laptop ecosystem, as reported by Wccftech.

Early Access to Rubin GPU Architecture

In addition to Arm support, CUDA Toolkit 13.4 provides early developer access to the NVIDIA Rubin architecture, designated with compute capability 107, as a preview release. According to NVIDIA, Rubin is designed as the next-generation GPU architecture powering the era of agentic AI. The preview enables application maintainers and developers to begin porting their software before general availability is introduced in a future version of the toolkit, according to official release notes. Wccftech reports that this early access is particularly critical for maintainers of key libraries to validate compatibility for Vera Rubin CPUs and GPUs.

Multi-Process Service V3 and Compute Fabric Updates

The 13.4 release introduces Multi-Process Service V3, bringing a modernized control layer that simplifies the automation and management of shared GPU resources, according to NVIDIA. MPS V3 provides a scriptable command-line interface, named server instances, and namespaces to organize concurrent workloads. It also adds TOML configuration support, streaming multiprocessor partition controls, and cgroup-integrated GPU memory limits. These additions allow for precise partitioning of compute performance, memory boundaries, and execution priority in containerized environments. Furthermore, the release debuts CUDA Compute Fabric Transport, a transport-centric API that lets advanced applications and communication libraries move data across the NVLink fabric at scale. Instead of mapping remote GPU allocations directly into a process’s virtual address space, software can target named logical endpoints using endpoint IDs and offsets. Wccftech points out that the update also expands CUDA Python with version 1.1.0, adding texture and surface programming, NUMA-aware managed memory, and type stubs.