Developers Get Windows 11 Running on an Intel NPU Instead of a GPU

Developers have managed to render the Windows 11 desktop using an Intel neural processing unit (NPU) instead of relying on a traditional GPU. However, the experiment comes with a major limitation: the desktop runs at just a few frames per second.

A developer group called labscommunity created an open-source project called npu-render, which uses a custom display driver to handle desktop rendering through an Intel NPU.

The developers successfully rendered several Windows 11 components and applications, including File Explorer, the Start menu, Notepad, and Microsoft Edge.

How developers managed to render Windows 11 using an NPU

The project relies on Intel’s OpenVINO framework and custom SHAVE kernels to process rendering workloads on the NPU.

Instead of modifying Windows applications, npu-render uses a custom display driver that redirects rendering tasks to the NPU. This allows applications to run normally without requiring changes to how they interact with Windows.

The prototype currently supports Intel NPUs only, and the developers have focused on proving that an NPU can handle desktop rendering rather than delivering practical performance.

Windows 11 runs at just 2–3 FPS on an Intel NPU

While the experiment successfully renders the Windows desktop, performance remains extremely limited.

Testing on an Intel Panther Lake laptop showed that npu-render could achieve only 2–3 screen updates per second, falling significantly short of the project’s 30Hz target.

One documented test recorded approximately 1.6 FPS, making even basic desktop interactions impractical.

However, the experiment also demonstrated that rendering workloads could shift away from the GPU. In one test, GPU 3D engine utilization dropped from 3.4% to 0.6%, while NPU utilization increased to approximately 58%.

Although these figures confirm that the NPU handled part of the rendering workload, the developers cautioned that the results do not indicate meaningful performance improvements for other GPU-intensive tasks.

Could NPUs eventually replace GPUs for some Windows tasks?

For now, npu-render remains a research prototype rather than a practical alternative to GPU rendering.

The experiment demonstrates that NPUs, which manufacturers primarily design to accelerate AI workloads, can also handle tasks outside traditional AI inference.

However, with performance limited to just a few frames per second, NPUs remain far from replacing GPUs for everyday Windows desktop rendering.

Still, the project offers an interesting demonstration of how developers could use dedicated AI hardware for workloads beyond its original purpose.

In other AI news, Outlook Search is getting smarter with Microsoft’s new AI model, while a recent report found that 84% of UK business leaders back AI, but workers fear job losses. Meanwhile, OpenAI’s data centers might soon be called AI compute campuses.

Via Neowin