Your next Windows computer upgrade will most likely have a new type of processor. It is called an NPU and is the most important requirement for Microsoft’s next generation of Windows computers\u2014Copilot+ PCs .\u00a0 While the new Snapdragon processors and Microsoft\u2019s requirements brought NPUs into the spotlight, they\u2019ve been around longer than you think. However, these new processors mean you no longer have to rely on cloud servers to run generative processes.\u00a0 So what are NPUs? How do they work? What are their benefits and drawbacks? I\u2019ll explain everything in this article.<\/p>\n
NPUs (short for neural processing units) are dedicated chips that execute complex AI operations. They process multiple data points in parallel and simulate brain neurons. This computing method is the ideal way to run neural networks and tackle functions like natural language processing, image recognition, and generative tasks.<\/p>\n
hey are faster than GPUs <\/a> and CPUs <\/a> for these and more AI use cases as a result.<\/p>\n As mentioned, NPUs are not as new as their current hype. Before the Microsoft Build event, they\u2019ve resided in CPUs, GPUs, smartphones, and servers. They\u2019ve also gone by different names, such as tensor cores, neural engines, and DSPs (digital signal processors).<\/p>\n For example:<\/p>\n NPUs have also evolved from what they used to be. This evolution was primarily driven by the ever-expanding list of AI use cases, especially generative AI.<\/p>\n Earlier versions focused on less complicated object identification use cases, which relied on simpler neural networks known as CNNs (convolutional neural networks) <\/a>. But AI video and photography increased the stakes. NPUs that could handle more advanced neural network models were developed as a result.<\/p>\n Generative AI, which involves using AI to generate new content, brought even higher power and performance demands.<\/p>\n Creating new text, image, video, and audio output requires more memory and data transfer, which could lay waste to limited bandwidth and spike energy consumption. Today\u2019s NPUs solve these problems for smaller devices as they are integrated into large systems on chips (SoC) <\/a>.<\/p>\n What most people call CPUs today are systems on chips or systems-on-a-chip. They are integrated circuits that contain different processing units, such as CPUs, GPUs, and TPMs. This chip design significantly saves space, improves power efficiency, and boosts performance.<\/p>\n The CPU has to hand off tasks to the GPU, TPU, or NPU when users initiate specific tasks. If these processors are separate, the CPU must transfer data over relatively long distances on the same board, which would consume more bandwidth and drive up energy usage. It means laptops, tablets, and phones will easily overheat <\/a> and slow down because they lack space.<\/p>\n Also read: <\/b>Ultimate Guide: How to Stop a Laptop from Overheating <\/a><\/p>\n That is why most gaming laptops with separate GPUs consume more power and run loud fans to dissipate excessive heat. Integrating these processors on a single chip resolves these issues and allows smaller form factors to perform faster and consume less energy.<\/p>\n We\u2019ve established that NPUs are better at handling complex AI tasks than other types of processors. But that doesn\u2019t mean they can replace them. CPUs and GPUs still have important computing roles to play. With the emergence of AI , NPUs should be seen as complementary processors. NPUs ensure AI features will not bog down your GPU and CPU. They will allow them to handle regular and demanding functions.\u00a0 On the flip side, NPUs are not versatile and cannot handle most of the processes CPUs and GPUs are designed to execute.<\/p>\n CPUs (central processing units) are commonly considered every computer\u2019s brain. They are general-purpose processors that deal with data manipulation and transfer, logical operations, input and output tasks, some graphics processing, and emulation. Heavy tasks like simulation and development rely mostly on CPUs. They also determine whether to send tasks to the GPU or NPU. While they can do most things, it\u2019s better to outsource specialized tasks to other processors. They are especially not optimized for parallel processing like NPUs. You\u2019ll notice significant slowdowns if you force machine learning and AI operations to run on the CPU.<\/p>\n GPUs are designed to handle demanding graphical computations. They\u2019re more valuable to gamers, video editors, and 3D modelers. That said, NVIDIA has made the argument that GPUs are great for AI tasks <\/a> because they have numerous integrated cores and excel at parallel processing. But their AI capabilities are mostly targeted at graphics-related tasks and features, such as DLSS and live streaming HDR. These tasks also mostly run on AI-optimized cores known as tensor cores. Dedicated NPUs do a better job at other use cases, such as system-wide information processing and other generative applications.<\/p>\n Microsoft\u2019s Copilot+ PC is mostly responsible for the current spotlight on NPUs. The company is boldly betting on AI to revolutionize Windows and transform how people interact with their computers. Its new Copilot+ PCs offer exciting new integrated and generative AI capabilities powered by dedicated NPUs. That way, GPUs and CPUs are not burdened by these new classes of features.<\/p>\n Copilot+ PC is Microsoft\u2019s new class of Windows computers. Devices with the Copilot+ brand ship with AI features that will not run on other computers. They include:<\/p>\n <\/div>\r\n<\/div>\r\n\n Do you need a Copilot+ PC<\/a>? Our rundown on these next generation computers will tell you all you need to know.<\/p>\n We\u2019ve seen teasers in the latest Windows 11 build, such as the preview version of the Copilot app <\/a> and Image Creator in Paint. However, Microsoft requires NPUs that can pull off 40 trillion operations per second (TOPS) to run the main Copilot+ AI features. Currently, only the new Snapdragon X processors can manage that capability. The new chipsets are behind Microsoft\u2019s claim that Copilot+ PCs are the \u201cfastest and most intelligent\u201d Windows devices ever. Also Read: <\/strong>Which Copilot+ PC Battery Offers the Best Performance: Let\u2019s Compare<\/a><\/p>\n For example, whenever you press the Copilot button, launch Recall, activate Live Captions, or turn on Cocreator, Windows will automatically route those processes to the NPU. You can also see the NPU\u2019s real-time graph in the Task Manager tick up to show that it\u2019s busy.<\/p>\n NPUs also handle other integrated AI functions, such as graphics upscaling (automatically improving video and image equality, especially in games).<\/p>\n Additionally, they free up computing resources for GPUs and CPUs by handling tasks related to AI that these processors may have had to handle in the past. These tasks include things like graphics upscaling and object recognition. That way, you\u2019ll notice significant performance improvements.<\/p>\n As mentioned, NPUs are not exclusive to computers. They\u2019ve also been integral to other smart devices for years.\n
How NPUs have evolved<\/h2>\n
What are SoCs, and why are they important?<\/h2>\n
\n
\nNPUs vs. Traditional Processors<\/h2>\n
\n
NPUs vs. CPUs<\/h3>\n<\/li>\n<\/ul>\n
\n
NPUs vs. GPUs<\/h3>\n<\/li>\n<\/ul>\n
\n
Windows and NPUs<\/h3>\n<\/li>\n<\/ul>\n
\n
\n
\n
\n
\nOther Types of Devices Using NPUs<\/h2>\n
\nMobile chip makers such as Qualcomm, Samsung, Apple, and Huawei have used NPUs for years to power features like voice assistants, facial recognition, predictive typing, translation, and task automation.<\/p>\n