What is an NPU? The AI engine inside new laptop processors

An NPU (neural processing unit) is the part of a modern laptop processor built specifically for AI tasks, such as blurring your background on a video call, generating live captions or cutting out background noise. It handles that work using far less power than the CPU or graphics chip. A strong NPU matters mainly if you want Windows’ Copilot+ features or better battery life during AI-heavy tasks; for cloud chatbots, it barely matters.

What an NPU actually does

AI features run on neural networks, and neural networks boil down to enormous numbers of simple multiply-and-add calculations. An NPU is circuitry designed to do exactly that math, many operations at once, at low power. Microsoft’s Copilot+ PCs developer guide describes it as processing large amounts of data in parallel while using energy on AI tasks more efficiently than a CPU or GPU, which helps battery life.

In almost every laptop, the NPU isn’t a separate chip or an add-on card. It is a block built into the main processor, next to the CPU cores and the graphics. You’ll find one in Intel Core Ultra, AMD Ryzen AI and Qualcomm Snapdragon X chips, and Apple calls its version the Neural Engine in M-series chips.

The three engines split the work:

Part	Best at	Typical AI role	Power draw for AI
CPU	Running Windows, macOS and everyday apps	Small or occasional AI jobs, and preparing data	Moderate, but inefficient for sustained AI
GPU	Graphics, games and heavy parallel math	Big bursts, such as image generation and large local models	High
NPU	Repetitive AI math at low precision	Long-running features such as video effects, captions and on-device search	Low

According to Microsoft, the NPU works alongside the CPU and GPU, and Windows 11 assigns each task to the most appropriate one. NPUs are fastest with lower-precision numbers, such as 8-bit integers (INT8), so AI models are usually converted, or “quantized,” before they run there.

Where you’ll find one in 2026

Here is how the main laptop chip families compare, using each maker’s own peak rating:

Chip family	AI engine	Maker’s NPU rating
Intel Core Ultra 200V (Lunar Lake)	Intel NPU	Up to 48 TOPS
Intel Core Ultra Series 3 (Panther Lake)	Intel NPU	Up to 50 TOPS
Intel Core Ultra 200H and 200HX	Older Intel NPU design	About 13 TOPS
AMD Ryzen AI 300 series	XDNA 2 NPU	50 to 55 TOPS
AMD Ryzen AI 400 series	XDNA 2 NPU	Up to 60 TOPS
Qualcomm Snapdragon X series (first generation)	Hexagon NPU	45 TOPS
Qualcomm Snapdragon X2 Elite	Hexagon NPU	80 to 85 TOPS
Apple M4	Neural Engine	38 trillion operations per second
Apple M5	16-core Neural Engine, plus Neural Accelerators in the GPU	No TOPS figure published

A few notes on those numbers. AMD says every Ryzen AI 400 chip exceeds the Copilot+ PC requirements. Microsoft puts the first Snapdragon X series at 45 TOPS, and Qualcomm’s Snapdragon X2 Elite product brief lists 80 or 85 TOPS depending on the model. Apple rated the M4 Neural Engine at 38 trillion operations per second, but for M5 it highlights a faster Neural Engine and a Neural Accelerator in each GPU core instead of a TOPS number. Makers don’t always measure the same way, so treat cross-brand comparisons with caution.

What TOPS means, and why it isn’t the whole story

TOPS stands for trillions of operations per second. Qualcomm’s guide to TOPS explains that it is calculated from the number of multiply-accumulate units in the NPU and their clock speed, and that the industry usually quotes it at INT8 precision. In other words, it is a theoretical peak, not a measured result.

Keep four caveats in mind:

• Peak, not typical. AMD’s own footnote describes its TOPS figure as the maximum possible “in an optimal scenario” that “may not be typical.”
• NPU TOPS versus platform TOPS. Some makers also advertise a “platform” or “total” TOPS figure that adds the CPU and GPU. The Copilot+ requirement counts the NPU alone.
• Software decides. Microsoft notes that the NPU needs software specifically programmed to use it. An app that ignores the NPU gains nothing from a bigger number.
• Memory matters too. Larger AI models depend on how much memory the laptop has and how quickly it moves data, not just raw TOPS.

A useful rule: treat 40 TOPS as a pass-or-fail line for Copilot+ features. Above that line, a higher rating gives headroom for future software, but it’s not a reliable guide to how fast a particular app will run.

What the NPU does on a laptop today

On Windows

• Video calls. Windows Studio Effects, which include background blur, eye contact correction, automatic framing and voice focus, require an NPU. Microsoft says NPUs below 40 TOPS get only some of the effects, and a few extras, such as portrait light and creative filters, are limited to Snapdragon Copilot+ PCs.
• Copilot+ PC features. On laptops with a 40+ TOPS NPU, Windows adds exclusive features such as Recall, which lets you search a timeline of screen snapshots by describing what you remember; Click to Do, for quick actions on text and images on screen; Live Captions that translate audio from more than 40 languages into English, even offline; and Cocreator image generation.
• Apps. Developers can use Phi Silica, a small language model that runs on the NPU of Copilot+ PCs, to add on-device text features; Microsoft is replacing it with Aion Instruct from late 2026.

Recall is opt-in, and Microsoft says its snapshots are stored and analyzed locally on the PC rather than sent to Microsoft.

On a Mac

Apple says the Neural Engine accelerates on-device AI tasks and the generative models behind Apple Intelligence, working alongside the CPU and GPU. Apple Intelligence runs on Macs with M1 or later, plus the A18 Pro MacBook Neo.

Check your own laptop

On Windows 11, open Task Manager and go to the Performance tab. If your laptop has a supported NPU, it shows up next to the CPU, memory and GPU, along with how busy it is.

What an NPU doesn’t do

• It doesn’t speed up cloud chatbots. When you use ChatGPT, Gemini or the Copilot app, most of the work happens on remote servers, so any laptop with a decent internet connection does the job.
• It doesn’t run your games. The GPU renders games. A bigger NPU won’t raise your frame rates.
• It doesn’t replace a strong GPU for heavy local AI. As of 2026, many popular tools for running large language models or image generators on your own machine still lean on the GPU and plenty of memory. NPU support in those tools is growing but uneven.
• It doesn’t train AI models. Laptop NPUs are built to run finished models, a job called inference, not to create new ones.

Copilot+ PC requirements, explained

Microsoft sets these minimum specs for a Copilot+ PC:

• An NPU capable of at least 40 TOPS
• 16 GB of RAM
• A 256 GB SSD
• Windows 11 version 24H2 or newer

Recall asks for a bit more. Microsoft’s Recall requirements add at least 8 logical processors, 50 GB of free space to turn it on, Device Encryption or BitLocker, and Windows Hello Enhanced Sign-in Security with a fingerprint or face sign-in.

The easiest check is the Copilot+ PC label on the box or product page. The phrase “AI PC” on its own is not a guarantee: first-generation Core Ultra laptops and those with Core Ultra 200H or 200HX chips, for example, have NPUs well under 40 TOPS.

Should an NPU change what you buy?

Prioritize a Copilot+ PC if you:

• want Recall, Click to Do or offline live caption translation
• spend hours on video calls and want effects that go easy on the battery
• keep laptops for several years and want room for on-device AI features as software catches up

Don’t pay extra for NPU speed if you:

• mostly use cloud chatbots, office apps and streaming
• are buying for gaming, where the GPU matters far more
• would have to give up 16 GB of RAM or a better screen to afford it

A quick checklist before you buy:

1. Want Windows’ on-device AI features? Look for the Copilot+ PC label or an NPU rated at 40 TOPS or more.
2. Get 16 GB of RAM at minimum, and 32 GB if you plan to run AI models locally.
3. On a Snapdragon laptop, confirm your must-have apps and accessories work on Windows on Arm.
4. On a Mac, every M1-or-later model and the MacBook Neo support Apple Intelligence, so choose based on memory, storage and screen.

The bottom line

An NPU is the power-efficient AI engine built into modern laptop processors. It runs video-call effects, live captions and Copilot+ features without draining the battery, but it doesn’t speed up cloud chatbots or games. If you want Windows’ on-device AI features, buy a Copilot+ PC; otherwise, treat the NPU as a bonus and spend on RAM, storage and a good screen. Our guide on how to choose a laptop covers those specs, the tech glossary defines the jargon, and you can find more coverage in AI and across all our guides.

You can find more plain-English AI explainers on RedAndWhiteMagz.com.