随着人工智能处理器逐渐普及，越来越多的PC开始配备专门用于AI计算的神经网络处理单元（NPU）。不过，这类硬件目前主要用于本地AI模型推理和图像处理，日常Windows桌面仍然主要依赖集成显卡或独立显卡完成图形渲染。
近日，一个名为labscommunity的开发团队推出了开源项目npu-driver，成功实现了一项颇为特殊的实验：让Windows 11完全通过NPU完成桌面图形合成，而不再依赖传统GPU。虽然目前运行速度极其缓慢，但这一成果证明，原本主要用于人工智能计算的NPU，也能够承担部分传统图形处理任务。
该项目采用Apache 2.0开源许可证，源代码已经公开发布在GitHub上。开发团队的目标是探索一个问题：能否在不修改Windows应用程序和桌面合成器的情况下，将原本交由GPU完成的桌面渲染工作，完整转移到NPU上？
为了实现这一目标，开发人员设计了一套特殊的显示驱动程序，截取Windows桌面窗口管理器（DWM）原本发送给显卡的图形操作，再将这些操作转换成NPU能够处理的计算任务。
Windows应用程序仍然按照传统方式运行，记事本、文件资源管理器、开始菜单、Microsoft Edge以及其他窗口并不需要针对NPU进行专门适配。真正发生变化的是底层图形处理路径：原本由GPU执行的桌面合成任务，被重新分配给了NPU。
在具体实现上，npu-driver采用自定义Direct3D 10用户模式驱动程序，将图形操作记录为一批批帧数据，再通过OpenVINO和专门的SHAVE计算内核交给Intel NPU处理。系统还实现了自己的光栅化处理模块，负责顶点处理、纹理采样、图层混合等传统图形操作。
目前，这套方案已经在搭载Intel Panther Lake处理器的笔记本电脑上完成测试。Panther Lake采用第五代NPU架构，相关处理器的NPU算力达到45至50 TOPS。
然而，实验结果也暴露出一个非常明显的问题：性能。
目前npu-driver的桌面刷新速度只有每秒2至3帧，部分测试甚至低至0.6帧。相比之下，普通Windows桌面通常以每秒60帧或更高的速度运行。即使不涉及游戏，仅仅移动窗口、打开文件资源管理器或者操作任务栏，也会变得异常缓慢。
换句话说，这套方案目前能够证明NPU可以承担完整的桌面图形合成任务，却远远达不到日常使用所需的性能水平。
开发人员的目标是将刷新速度提高到每秒30帧，但目前距离这一目标仍有相当大的差距。
为了避免实验过程中出现严重故障，开发团队还设计了自动回退机制。当NPU驱动发生错误，或者用户主动要求切换时，系统可以重新交由集成显卡或独立显卡完成图形渲染。不过，目前这一恢复机制仍然存在一些限制，部分测试中并不能完整恢复所有应用程序窗口的位置。
项目开发者特别强调，npu-driver目前只是研究原型，而不是面向消费者的成熟产品。它需要特定的Intel Panther Lake硬件、兼容的NPU驱动程序以及额外的开发工具，普通用户并不能简单安装后就获得无GPU运行的Windows 11系统。
这项实验的意义，实际上并不是要证明NPU可以取代显卡，而是探索如何重新分配现代处理器内部不同计算单元的工作。
如今的CPU、GPU和NPU各自拥有不同的设计目标。GPU擅长大规模并行图形计算，NPU则针对矩阵运算和神经网络推理进行了专门优化。让NPU承担桌面渲染任务，意味着开发者必须将传统图形处理流程重新转换为适合神经网络计算单元执行的任务，这本身就是一项相当复杂的工程。
不过，如果未来能够显著提高处理效率，NPU或许可以承担部分原本由GPU负责的轻量级桌面任务，从而释放GPU资源用于游戏、视频处理或本地AI推理。
目前，npu-driver距离真正实用还有很长的路要走，但它已经完成了一个有趣的技术验证：Windows 11的桌面并非只能通过传统GPU进行合成渲染，现代处理器中的NPU同样能够接管这项工作，只是现阶段付出的性能代价实在过于巨大。
了解更多：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
As artificial intelligence processors become widespread, more and more PCs are equipped with neuronet treatment units (NPUs) dedicated to AI calculations. However, this type of hardware is currently used mainly for local AI model reasoning and image processing, and the daily Windows desktop continues to rely mainly on integrated GPU or stand-alone GPU for graphic rendering.
In recent days, a development team called LabsCommunity launched the Open Source project npu-driver, which succeeded in achieving a rather special experiment: allowing Windows 11 to complete desktop graphic synthesis exclusively through NPU rather than relying on traditional GPU. Although currently operating at a very slow pace, this result proved that NPUs, which were used mainly for artificial intelligence calculations, were also able to undertake some of the traditional graphic processing tasks.
The project uses the Apache 2.0 open source licence, and the source code has been made public on GitHub. The objective of the development team is to explore the question of whether, without modifying Windows applications and desktop synthesizers, the desktop rendering originally assigned to GPU could be fully transferred to NPU.
To achieve this goal, the developer designed a special display drive to intercept the graphical operation that Windows Desktop Window Manager (DWM) had originally sent to GPU and then convert it to a calculation task that NPU could handle.
Windows applications are still operating in the traditional way, and notebooks, file resource managers, start-up menus, Microsoft Edge and other windows do not need to be tailored to NPU. The real change is the bottom graphic processing path: the desktop synthesis task originally performed by GPU was redistributed to NPU.
In practical terms, npu-driver uses the self-defined Direct3D 10 user mode driver to record graphic operations as batches of frame data, and then calculates kernels to Intel NPU through OpenVINO and SHAVE. The system has also achieved its own rasterization module, which is responsible for traditional graphic operations such as vertex processing, texture sampling and layer mixing.
The program has now been tested on the laptop that carries the Intel Panther Lake processor. Panther Lake uses the 5th generation NPU structure, with NPU calculators ranging from 45 to 50 TOPS.
However, the results of the experiment also revealed a very obvious problem: performance.
The current desktop update rate for npu-driver is only 2 to 3 frames per second, and some tests are even as low as 0.6 frames. In contrast, regular Windows desktops usually run at 60 frames per second or more. Even if the game is not involved, simply moving the window, opening the file resource manager or operating the taskbar becomes extremely slow.
In other words, the scheme currently proves that NPU can perform full desktop graphic synthesis tasks, but far below the performance level required for daily use.
The goal of the developers is to increase the rate of retrofitting to 30 frames per second, but there is still a considerable gap from this target.
In order to avoid a serious malfunction in the course of the experiment, the development team also designed the automatic rollback mechanism. When an error occurs in the NPU driver, or when the user voluntarily requests a switch, the system can revert to the integrated GPU or independent GPU to finish the graphic rendering. However, there are still some limitations to this recovery mechanism and some of the tests do not fully restore the location of all application windows.
In particular, project developers emphasized that npu-driver is currently only studying prototypes, not mature products that are consumer-oriented. It requires specific Intel Panther Lake hardware, compatible NPU drivers and additional development tools, and ordinary users do not simply install and acquire Windows 11 systems that are not run by GPU.
The purpose of the experiment was not to prove that NPU could replace GPU, but to explore the reallocation of different computing units within modern processors.
Today, CPUs, GPUs and NPUs have different design objectives. GPU specializes in large-scale parallel graphic calculations, while NPU specializes in matrix computing and neuronet reasoning. Making the NPU a desktop rendering task means that the developers have to convert the traditional graphic processing process to a task suitable for the neuronet computing module, which is itself a rather complex project.
However, if processing efficiency could be significantly improved in the future, NPU might be able to undertake some of the lightweight desktop tasks that were originally the responsibility of GPU, thereby releasing GPU resources for games, video processing or local AI reasoning.
At present, npu-driver has a long way to go from being really practical, but it has completed an interesting technical test: Windows 11's desktop is not only synthesized through traditional GPUs, and NPUs in modern processors are equally capable of taking over the job, but the performance costs are so great at this stage.
Learn more:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.