What Is an NPU, and Do You Need One in a Laptop?

What an NPU does, next to the CPU and GPU

A modern laptop processor is really three kinds of engine on one package. The CPU runs the operating system and general programs and is good at a wide mix of tasks. The GPU draws the screen and is good at doing the same math on huge amounts of data in parallel, which is why it also runs games and most AI software. The NPU is the newest block: a small accelerator designed only for the math inside AI models.

Microsoft's Copilot+ PC developer guide describes the NPU as a specialized chip for AI-intensive processes that processes large amounts of data in parallel and uses energy on AI tasks more efficiently than a CPU or GPU, which results in longer battery life. The same guide says the NPU works alongside the CPU and GPU, and that Windows 11 assigns each task to the most appropriate place.

The practical effect is simple. A feature that runs all the time in the background, such as blurring your webcam background or captioning audio, can sit on the NPU and leave the CPU and GPU free for your actual work, without draining the battery as fast as it would on the GPU.

What TOPS means and how makers use it

TOPS stands for trillion operations per second. It is the number chip makers use to rate AI throughput, and it is always a peak figure: the most operations the hardware can do in ideal conditions, not what a given app will see.

Two details make TOPS figures easy to misread. First, the number depends on the data format. Intel labels its figures as INT8, meaning 8-bit integer math, and Microsoft's guide notes that many NPUs only support lower-bit integer math such as INT8, so models trained in formats like FP32 have to be converted (quantized) before they run on the NPU. Second, makers quote TOPS for different parts of the chip. Intel's spec pages list separate NPU, GPU and overall peak TOPS; the overall number covers the whole processor, not the NPU alone.

So when comparing laptops, compare NPU TOPS with NPU TOPS. The overall platform number is useful as a ceiling, but only software written for all three engines can approach it, and Windows features that require an NPU look only at the NPU.

NPU figures for current laptop chips

The table below lists NPU figures as the chip makers or their laptop partners state them on official spec pages. Microsoft's threshold for Copilot+ PCs is an NPU that can perform more than 40 TOPS, and its guide names Snapdragon X Elite, AMD Ryzen AI 300 series and Intel Core Ultra 200V series as qualifying chip families.

Note the Intel Core Ultra 7 255H: it has an NPU, but at 13 TOPS it sits well below the 40 TOPS line, while its GPU is rated far higher. A chip having an NPU does not by itself make a laptop a Copilot+ PC. For how the Intel names map to these parts, see Intel Core Ultra names explained.

Apple takes a different approach. Its announcement of the M5 chip describes a 16-core Neural Engine (Apple's name for its NPU) plus a Neural Accelerator in each of the 10 GPU cores, but gives no TOPS figure for either. On a Mac, Apple Intelligence and apps built on frameworks such as Core ML use this hardware automatically, so there is no Copilot+ style threshold to check.

What software uses the NPU today

An NPU does nothing unless software is written for it. Microsoft's developer guide says it plainly: like other hardware, the NPU needs software specifically programmed to take advantage of it. Today that software falls into three groups.

The first is Windows itself. Microsoft's announcement of Copilot+ PCs on AMD and Intel chips lists the experiences these PCs get: Live Captions with translation, Windows Studio Effects (background blur, portrait blur, eye contact and automatic framing), Cocreator in Paint, Restyle image and Image Creator in Photos, and Recall. The second is apps that call the Windows AI APIs. Microsoft Learn states that on a Copilot+ PC these APIs always run on the NPU, and lists Phi Silica (a small on-device language model), text recognition, speech recognition, video and image super resolution, image description, image segmentation, Object Erase and image generation.

The third group is apps that bring their own models through Windows ML, which Microsoft now recommends for running AI on the NPU. Windows ML picks the best execution provider on the device (for example Qualcomm's QNN for Snapdragon NPUs or Intel's OpenVINO for Intel NPUs) and falls back to the GPU or CPU when needed. You can see whether anything is using the NPU in Task Manager, which now shows NPU usage next to CPU and GPU.

• Always-on Windows features: Studio Effects, Live Captions, Recall and the Copilot+ experiences.
• Windows AI APIs that apps call: Phi Silica, OCR, speech recognition, imaging and image generation.
• Third-party apps that target the NPU through Windows ML and ONNX Runtime.

When a GPU is still the better tool for AI

The NPU is built for efficiency, not raw size. It runs small models that were quantized and packaged for it. If you want to run open language models of your choice, generate images with community tools, or train and fine-tune models, a GPU is still the tool that matters, and the amount of GPU memory decides which models load at all. Our guide to laptop specs for local AI explains how to size that memory.

Intel's own numbers show the gap on a single chip: the Core Ultra 7 255H lists 74 TOPS for its GPU against 13 TOPS for its NPU. Microsoft's API table points the same way from the software side. Phi Silica runs on the NPU of a Copilot+ PC, and Microsoft is extending it to GPUs, listing NVIDIA GeForce RTX 30 series and newer and AMD Radeon RX 9060 series and newer with at least 6GB of video memory, but that GPU path currently requires Developer Mode.

Most open-source AI tooling is also built for GPUs first. If local AI is your main reason to buy, start from the laptops for AI guide, which ranks picks by GPU memory, and treat the NPU as a bonus for battery life.

Who needs an NPU in a laptop

You will get real use from a 40+ TOPS NPU if you spend a lot of time on video calls and want Windows Studio Effects without a battery hit, if you want Live Captions with translation, Recall or the image features in Photos and Paint, or if your company plans to deploy apps built on the Windows AI APIs. For these users the NPU is the deciding spec, and our explainer on what a Copilot+ PC is covers the rest of the requirements.

Business buyers have an extra reason to look at the NPU: a laptop bought today will likely stay in service for several years, while more Windows and third-party features move to on-device AI. Choosing a Copilot+ class chip is a way to avoid ruling those features out. The business laptops guide is the place to start.

You can skip the NPU question if you mostly browse, write and stream and do not care about the Copilot+ features, if you use a Mac (every Apple silicon chip has a Neural Engine anyway), or if your AI work is running your own models, where the GPU and memory matter far more.

What to pick for your use

Questions

Is an NPU the same as a GPU?

No. Both do parallel math, but a GPU is a general graphics and compute engine that runs games, video and most AI software, while an NPU is a smaller accelerator designed only for AI model math at low power. Microsoft describes the NPU as more energy efficient than a CPU or GPU for AI tasks.

How many TOPS does a laptop NPU need?

For Copilot+ PC features, Microsoft requires an NPU that can perform more than 40 TOPS. Below that, the NPU can still accelerate apps that target it, but you will not get the Copilot+ experiences.

Does a higher TOPS number mean faster AI?

Only roughly. TOPS is a peak figure measured at a stated data format, usually INT8, and real speed depends on how well the software and model are optimized for that NPU. Compare NPU TOPS with NPU TOPS, not with a chip's overall platform figure.

Can I run ChatGPT-style models on the NPU?

Small ones, yes. Windows ships Phi Silica, a small language model that runs on the NPU of Copilot+ PCs, and apps can use it through the Windows AI APIs. Larger open models usually run on a GPU. See our guide to local AI laptop specs.

Do MacBooks have an NPU?

Yes. Apple calls it the Neural Engine. The M5 has a 16-core Neural Engine, and Apple says it speeds up Apple Intelligence and apps that use frameworks such as Core ML. Apple does not publish a TOPS figure for it in its M5 announcement.

How can I tell if my laptop is using its NPU?

On Windows 11, open Task Manager and look at the Performance tab. Microsoft says devices with an NPU show its usage there next to the CPU and GPU.