NVIDIA releases first CUDA Toolkit preview for Windows on Arm and RTX Spark

VideoCardz reports that NVIDIA has released the CUDA Toolkit 13.4 Developer Preview, marking the first official CUDA development package for Windows on Arm64. The release is specifically aimed at developers building applications for the upcoming RTX Spark platform — NVIDIA's Arm-based superchip combining a custom 20-core Grace CPU with an RTX 5070-class Blackwell GPU and over 1,000 TOPS of AI acceleration. While NVIDIA has supported Arm64 devices through its Jetson embedded platform and Linux Arm64 SBSA packages for years, this is the first time a native Windows Arm64 CUDA Toolkit has been made available — a critical milestone for RTX Spark's software ecosystem ahead of the platform's consumer launch this fall.

CUDA Toolkit 13.4 includes full compiler and build-system support for Windows Arm64 targets, allowing developers to compile CUDA applications directly on Windows on Arm systems or cross-compile using the Windows x86-64 version of the toolkit. NVIDIA is providing local Windows Arm64 installer and binary archives containing CUDA headers, import libraries, runtime components, and all libraries required for compilation and linking. Importantly, NVIDIA notes that RTX Spark hardware itself is not required for porting — developers can begin optimizing their applications using any existing Windows on Arm system, with NVIDIA recommending that teams check third-party dependencies for Arm64 support and choose between native Arm64 and Arm64EC builds before validating on RTX Spark hardware. This lower barrier to entry is designed to accelerate software ecosystem readiness ahead of the first RTX Spark PCs shipping from Microsoft, ASUS, Dell, HP, Lenovo, and MSI.

The developer preview is clearly marked as pre-release software not intended for production deployment or benchmarking, and several known issues remain. CUDA transfers using pageable memory may perform below expectations, some PyTorch build workflows may trigger GPU timeouts or system restarts, the display may remain blank for approximately two minutes during driver installation, and Nsight Copilot is not available on Windows Arm64 in this release. NVIDIA has not announced a timeline for when Windows Arm64 CUDA support will move to general availability, just as RTX Spark itself lacks an official consumer release date beyond the Fall 2026 window. Nevertheless, the CUDA Toolkit 13.4 preview represents a concrete step toward making RTX Spark a viable development and deployment platform for AI and GPU-accelerated workloads on Windows on Arm, signaling that NVIDIA is actively investing in the tools and infrastructure developers need to build for its next-generation PC platform.