RTX Spark, NVIDIA launches the porting phase for developers ahead of the autumn release

NVIDIA has started distributing the first development drivers designed for the RTX Spark platform, the superchip destined for PCs with AI capabilities that will debut in the market during autumn 2026. This is a significant step for the Windows on Arm ecosystem, which thus receives its first official native CUDA package for this architecture.

RTX Spark is the chip that combines, in a single package, a processor based on the Grace architecture and a Blackwell GPU, accompanied by 128 GB of shared memory and connected via high-speed interconnect. The project is specifically designed for workloads related to artificial intelligence and will find a place on a wide range of notebooks and mini PCs coming by the end of the year.

The first driver package made available to developers is numbered 616.00 and is particularly aimed at the Microsoft Surface RTX Dev Box, a development system characterized by a completely fanless design. At the same time, NVIDIA has released the CUDA Toolkit 13.4 in Developer Preview version, which introduces native support for compilation for Windows on Arm64 for the first time.

It is worth noting that CUDA support for Arm systems is not entirely new: NVIDIA has long offered it on Jetson platforms and through Linux Arm64 SBSA packages. What was missing, until now, was a native CUDA toolkit for Windows Arm64, which this preview indeed fills.

The package allows for direct compilation on a Windows Arm64 system, as well as cross-compilation from a Windows x86_64 toolkit. Starting from this version, the driver is no longer included within the toolkit but is distributed separately through the NVIDIA Developer Driver channel, beginning with the R616 series. The preview provides headers, import libraries, and necessary runtime components for compilation and linking, while drivers, platform firmware, and operating system images remain standalone packages.

According to NVIDIA, developers can begin porting their applications to Windows on Arm right away using an existing development system, without necessarily having RTX Spark hardware. The recommended path includes checking the Arm64 compatibility of third-party dependencies, choosing between an Arm64 or Arm64EC porting strategy, compiling and testing the application on Windows on Arm, validating NVIDIA and CUDA software paths, testing installation, updates, features, and performance, and finally validating directly on RTX Spark hardware as soon as it becomes available. As this is a development version, the package has some issues already documented by NVIDIA.