What Is an NPU? Do You Need One in a PC? The Meaning of 40 TOPS and How It Is Used for AI

An NPU (Neural Processing Unit) is a dedicated chip inside a computer that handles AI calculations with less power. Many recent laptops include one. However, AI services such as ChatGPT, Claude, and Gemini run their calculations on servers, so the speed does not change whether or not you have an NPU. An NPU is useful when you use features that run AI on your own computer, such as the Copilot+ PC features in Windows or background blur in video calls.

This article explains, based on official materials from Microsoft, Intel, AMD, Qualcomm, Apple, and others (checked October 25, 2026), what an NPU is, which features use the NPU, how to read the performance figure (TOPS), and how to decide whether to choose a PC with one. For Macs, see Chapter 6. In this article, parts are color-coded: NPU GPU Cloud.

1. What Is an NPU? Differences from CPUs and GPUs

There are three parts in a computer that can perform AI calculations. AI calculations are centered on performing large numbers of multiplications and additions at the same time. Any of these parts can do the calculations, but each has different strengths and power use.

The name differs by manufacturer (Intel, AMD, and Qualcomm call it an NPU; Apple calls it the Neural Engine). In recent PC chips, the CPU, GPU, and NPU are packed into a single chip.

The key point is that an NPU only works when the app is built to use it. Microsoft's developer materials also state that an app must support the NPU in order to use it. Even if a PC has an NPU, apps that do not support it run on the CPU or GPU. In Windows 11, the NPU usage appears under "Performance" in Task Manager, so you can check whether the NPU is running while you use AI features (Microsoft "Copilot+ PCs developer guide").

2. Do ChatGPT, Claude, and Gemini Use the NPU?

The ChatGPT, Claude, and Gemini apps and web versions send your questions to servers on the internet, where large models generate the answers. The NPU and GPU in your PC are not used for generating answers, so they are not part of the operating requirements. For example, the operating requirements for the ChatGPT Windows app only concern the Windows version (OpenAI Help). Microsoft 365 Copilot also works by sending language model requests to Microsoft's data centers (Microsoft Learn).

So even if ChatGPT or Claude feels slow, switching to a PC with an NPU will not make it faster. The same applies to Claude Code and Codex. What your PC's performance changes is the time for local commands, such as builds and tests. For details, see "PC Specs Needed for Claude Code and Codex."

3. Windows Features That Use the NPU (Copilot+ PCs and More)

The NPU's effect is clearest in the features of Copilot+ PCs on Windows 11. A Copilot+ PC is the name for a PC that meets conditions set by Microsoft. The conditions are as follows (Microsoft "Windows 11 specifications and system requirements").

TOPS is the number of operations per second, in trillions. For how to read it, see Chapter 5. Microsoft states that PCs with an NPU of 40 TOPS or more but less than 16 GB of memory are not Copilot+ PCs. Even in that case, apps that support the NPU can still use it.

The main features available only on Copilot+ PCs are listed below, based on Microsoft's list. "Requires internet" means the feature requires a connection to Microsoft's cloud and an account for safety checks.

Source: Microsoft "Windows 11 specifications and system requirements," the "list of all Copilot+ PC features" (checked October 25, 2026). Availability varies by region and device.

Windows Studio Effects starts at 10 TOPS

Windows Studio Effects, which applies effects to camera video (background blur, eye contact, automatic framing, and voice focus), works even without a Copilot+ PC, as long as you have an NPU of 10 TOPS or more and a camera. Additional effects such as Portrait Light are only available on Copilot+ PCs with Snapdragon X (same Microsoft specifications page). The NPU handles the processing for these effects.

A small AI model built into Windows

Copilot+ PCs come with Microsoft's small language model Phi Silica preinstalled, and it runs on the NPU. It is the model that Windows features and apps use to summarize and rewrite text on your PC without connecting to the internet. Microsoft's materials compare running it on the NPU with running it on the GPU, saying the NPU responds faster and uses less power. Microsoft plans to replace Phi Silica with its successor model, "Aion Instruct," and for general PCs the plan is January 2027 (Microsoft Learn "Get started with Phi Silica").

4. Does a Local LLM Run on the NPU?

For "local LLMs," where you download an AI model to your own computer and run it, the GPU is still the main component. Many widely used tools are built to run on the GPU, and only some tools can use the NPU. The official support status is as follows (checked October 25, 2026).

Sources: Ollama "Hardware support", LM Studio "System Requirements", llama.cpp "Snapdragon", Microsoft Learn "Foundry Local architecture", AMD "Ryzen AI Software: LLMs", OpenVINO "Inference with GenAI on NPU".

Note that some AI that runs on your PC does not require an NPU. The Chrome browser includes a small Gemini model (Gemini Nano) that runs on your PC. However, the conditions Google gives are "a GPU with more than 4 GB of VRAM" or "a CPU with 16 GB or more of memory and 4 or more cores," and the NPU is not part of the conditions (Chrome for Developers).

When combining an NPU with an integrated GPU, AMD explains that the NPU handles the first stage, reading the input, and the GPU handles the second stage, generating the text. The division of labor is that the NPU suits tasks with a large amount of computation, while the GPU is better for tasks that keep reading memory.

If you want to run a local LLM, what you should look at is not the NPU's TOPS but the GPU's memory (VRAM) and the amount of computer memory. For a guideline, see "Specs Needed for Local LLMs," and for how to get started, see "How to Run a Local LLM."

5. Reading the Performance Figure: TOPS

NPU performance is expressed in TOPS (Tera Operations Per Second). It is the number of trillions of operations per second, so "50 TOPS" means 50 trillion operations per second. The chart below lists the NPU values of major chips, using the manufacturers' published figures (the maximum for each chip model). The red dashed line marks the 40 TOPS requirement for Copilot+ PCs. Each manufacturer measures precision differently, so use these figures only as a rough size guide (see note 2 below).

Sources: Qualcomm (the "Hexagon NPU TOPS (INT8)" row in the product materials), AMD (CES 2026 announcement), Intel (CES 2026 announcement).

• All of these are the maximum values within each chip family, and lower models are smaller (Snapdragon X2 Elite is 80, and Ryzen AI 400 is 50 to 60 depending on the model).
• To check whether a PC is covered as a Copilot+ PC, the most reliable way is to compare the names in Microsoft's list (Intel is written as "200V and 300V series," which differs from Intel's own "series 3" naming) with the "Copilot+ PC" label in the PC's spec sheet. Qualcomm's materials state that Snapdragon X2 Elite is intended for Copilot+ PCs.
• Earlier generations listed as covered Copilot+ PCs (Ryzen AI 300, Core Ultra 200V, Snapdragon X) also have NPUs of 40 TOPS or more, as required. Apple's M4 has 38 TOPS (Chapter 6).

Three cautions when comparing TOPS

1. Is it the NPU-only number or the whole-chip number? Intel and AMD also publish "whole-platform TOPS" that add the CPU, GPU, and NPU together. For example, Intel's desktop Core Ultra 200S has up to 13 for the NPU and up to 36 for the whole platform (Intel "Core Ultra 200S Product Brief"). Large numbers in advertisements may be whole-chip values, so when comparing against the Copilot+ PC requirement, check the NPU value.
2. The precision of the calculations is not consistent. Qualcomm states that its NPU value was measured in INT8 (8-bit integers), but AMD's and Intel's announcements do not state precision. NVIDIA's GPU "AI TOPS" figures use the coarser FP4 format, and they are values when using a feature that skips zero values in calculations (sparsity). In NVIDIA's materials, the desktop RTX 5090 shows two values, "1676/3352," and the larger one is the value with this feature used (NVIDIA RTX Blackwell technical materials). The AI TOPS of laptop RTX 50 series GPUs range from 440 to 1,824 (NVIDIA), but you cannot say they are "X times" the NPU figure.
3. TOPS is a theoretical maximum. Intel's same materials explain that TOPS is "the theoretical maximum when running at 100% efficiency." Actual speed depends on whether the app supports the NPU, memory speed, and power settings.

6. The Mac Neural Engine

Apple silicon Macs (M1 and later) include the Neural Engine, which corresponds to the NPU. Apple announced that the M4's Neural Engine performs 38 trillion operations per second, but it has not published TOPS figures for M5 and later. For M5, Apple added AI acceleration to each GPU core and explains the gains in AI performance mainly in terms of the GPU (Apple "M5" announcement).

For using AI features on a Mac, the guide is the list of supported models, not TOPS. Apple Intelligence works on Macs with M1 or later; the supported languages are listed by Apple (Apple Support "Apple Intelligence availability"). If you have an M1 or later Mac, you do not need to worry about the NPU performance thresholds that apply to Copilot+ PCs on Windows.

When running a local LLM on a Mac, the GPU is also the main component, not the Neural Engine. The Mac support listed in Ollama's materials is Apple's GPU (Metal), and Microsoft's Foundry Local also uses the GPU on Mac (Microsoft Learn). Because a Mac's GPU uses the computer's memory (unified memory), the size of the model you can run is determined by the amount of memory.

7. Should You Choose a PC with an NPU? Decision Steps

The main laptop chips that Intel, AMD, and Qualcomm released in 2026 all include an NPU. The decisions you need to make are "Should I choose a model that meets the Copilot+ PC requirements?" and "Should I spend money on the NPU or on the GPU?" Thinking through the following steps in order will help you decide.

1. Do you plan to use AI that runs on your own PC?
   No (only cloud services such as chat or Claude Code) The NPU is not a factor in your decision. You can choose based on memory, SSD, screen, and weight

   Yes Go to the next step
2. Will you choose a Mac?
   Yes With M1 or later, Apple Intelligence is available, and there is no TOPS threshold. (If you also want to run local LLMs, prioritize the amount of unified memory: Chapter 6)

   No (Windows) Go to the next step
3. Do you want to use Windows AI features?
   You also want Recall or Live Captions translation Choose a model labeled "Copilot+ PC" (NPU of 40 TOPS or more, 16 GB memory, 256 GB SSD) → Go to the next step

   Only Studio Effects such as background blur An NPU of 10 TOPS or more is enough → Go to the next step

   I won't use them The NPU is not a factor in your decision → Go to the next step
4. Do you want to run AI models on your own PC?
   Yes Prioritize the GPU's VRAM and the amount of memory over the NPU (Chapter 4)

   No You do not need to spend extra on a GPU for AI

Before buying, check the "Copilot+ PC" label and the chip model in the PC maker's spec sheet, and confirm the NPU TOPS on the chip maker's product page. For a PC you already have or a demo unit in a store, you can tell whether it has an NPU by checking whether Task Manager's "Performance" tab in Windows 11 shows an NPU entry. To see whether Studio Effects is available, go to Settings > Bluetooth & devices > Camera, select your camera, and check whether Studio Effects options appear (Microsoft's support page notes that if they do not appear, the PC may not have an NPU: Microsoft Support).

Summary

FAQ

Q. Can I use AI on a PC without an NPU?

Yes. ChatGPT, Claude, Gemini, and similar services run on servers, so you can use them the same way on a PC without an NPU. What requires an NPU is some features that are designed to run on your PC, such as the Copilot+ PC features in Windows.

Q. Is a higher NPU TOPS figure always better?

Within the same manufacturer and the same counting method, it is a useful guide. However, if the app you use does not support the NPU, the number has no effect. If you want to use Copilot+ PC features, it is enough to check whether the PC is labeled "Copilot+ PC" rather than looking at the TOPS figure (it must meet all of: NPU of 40 TOPS or more, 16 GB memory, and 256 GB SSD).

Q. For AI, which is better: a PC with a gaming GPU or a PC with an NPU?

For running heavy AI locally, such as local LLMs or image generation, a GPU with more VRAM is stronger. The NPU suits running light AI, such as background blur or transcription, continuously while keeping battery drain low.

Q. Does a Mac have an NPU too?

Yes. The Neural Engine in Apple silicon (M1 and later) corresponds to the NPU. On a Mac, it is simpler to think in terms of supported models rather than TOPS. Apple Intelligence works on Macs with M1 or later.

Q. Can I use Copilot without a Copilot+ PC?

Yes. The Copilot app and Microsoft 365 Copilot run in the cloud, so you can use them without a Copilot+ PC. A Copilot+ PC is required for features that run AI on your PC, such as Recall and Improved Windows Search.