On September 9th, NVIDIA unveiled the CUDA Toolkit 13.4, marking a significant milestone by offering, for the first time, a native development package tailored for the Windows on Arm64 platform. This move expands Arm platform support beyond Linux to encompass Windows on Arm as well. Developers are now empowered to compile CUDA applications directly on this system or undertake cross-compilation via the Windows x86-64 version.

The preview release also brings functional support for the NVIDIA Rubin architecture, with updates spanning compilers, programming interfaces, and software libraries. Notably, enhancements have been made to CUDA Python, Multi-Process Service V3, and CUDA Compute Fabric Transport.

The industry widely perceives this update as a preparatory step for the imminent launch of the first consumer-grade PC processor, RTX Spark, slated for October. The RTX Spark processor seamlessly integrates a Grace CPU with a Blackwell GPU, facilitating the local execution of large models boasting 120 billion parameters.

While this update does not directly boost the performance of existing devices, it establishes a robust foundation for CUDA application compatibility on the Windows on Arm platform. Once developers finalize the porting and validation of applications, RTX Spark users can anticipate a more extensive array of GPU-accelerated software support.