NVIDIA Releases First RTX Spark Developer Drivers Ahead of Fall Launch

NVIDIA is moving its RTX Spark platform closer to reality. The company released the first developer preview drivers this week—version 616.00—giving programmers their first real chance to prepare applications for hardware that won't ship until fall. The RTX Spark itself is a single-chip system combining NVIDIA's Grace CPU architecture, a Blackwell GPU, and 128 gigabytes of memory, all wired together on one piece of silicon designed specifically for AI workloads. It's coming to laptops and compact desktop machines, and NVIDIA is now making sure developers don't get caught flat-footed when those machines arrive.

The preview driver targets the Microsoft Surface RTX Dev Box, a fanless machine built explicitly for developers who want to start working with the platform before it reaches consumers. Alongside the driver release, CUDA 13.4—NVIDIA's toolkit for GPU programming—now includes what the company calls "Windows on Arm" preview support. This is the piece that matters most for the developer community. The toolkit enables two critical workflows: native development directly on Windows Arm64 systems, and cross-compilation, which lets programmers build Arm64 applications from existing x86_64 machines without needing the actual hardware yet.

The timing is deliberate. NVIDIA is essentially asking developers to start now. Review your code and your dependencies for Arm64 compatibility. Decide whether you're porting your application wholesale to Arm64 or using a hybrid approach called Arm64EC. Build and test on Windows on Arm systems if you can access them. Validate that your NVIDIA and CUDA software paths work as expected. Test installation, updates, functionality, and performance. Then, when RTX Spark hardware becomes available in the fall, validate everything one final time on the actual device. It's a methodical roadmap, and it assumes developers will take months to prepare.

But the preview is not without friction. NVIDIA documented several known issues that developers need to understand. CUDA transfers using pageable memory—the default kind—are delivering lower performance than expected. The workaround is to use page-locked memory instead, allocated through specific CUDA functions. During driver installation, the display may go blank for roughly two minutes; the company warns users not to restart or power off the system during this window, even though it looks like something has gone wrong. More concerning is a possible system instability when running PyTorch CI/CD workflows, which can trigger GPU timeouts and cause unexpected restarts or unresponsiveness. NVIDIA says this is under investigation. Nsight Copilot, a development tool, is also disabled in this preview and not available on Windows Arm64 yet.

These are the kinds of rough edges you expect in a developer preview. They're not showstoppers—they're signals that the platform is still being refined. NVIDIA has already shown RTX Spark systems at Computex 2026, and the company reported seeing impressive performance in AI applications, content creation tools, and games. This is NVIDIA's first serious push into the laptop and AI PC market with its own integrated system, and early impressions suggest the hardware is capable. But benchmarks from engineering samples have leaked online, and NVIDIA is explicitly cautioning against reading too much into them. Those numbers don't reflect what retail systems will actually deliver. The real test comes in the fall, when RTX Spark machines ship to customers and developers can measure performance on production hardware. Until then, the work is preparation—getting code ready, dependencies sorted, and workflows validated so that when the hardware arrives, applications are waiting.