NPU Laptops Explained: What an AI PC Really Does

What an NPU is

A laptop processor now contains three kinds of engine. The CPU is the general-purpose part that runs the operating system and most applications. The GPU draws the screen and is very fast at parallel maths, but draws a lot of power doing it. The NPU is a smaller block designed only for the multiply-and-add work at the heart of neural networks, at low precision and with very little power. Microsoft's developer guide describes the NPU as working alongside the CPU and GPU, with Windows assigning each task to the most suitable engine, and stresses the battery benefit of running AI work there instead of on the other two.

What TOPS means, and why it misleads

TOPS stands for trillions of operations per second, and it describes peak throughput at low precision, usually 8-bit integers. It is a ceiling, not a measure of how fast a real feature runs. Two chips with the same TOPS can behave very differently because of memory speed, software support and how long they sustain that rate before heat limits them. Microsoft's hardware requirements page sets the bar for its premium AI PC tier at an NPU capable of 40 or more TOPS, plus 16 GB of memory and a 256 GB SSD, and lists the processor families that currently qualify. Below that tier, a laptop may still have an NPU that powers a smaller set of features.

What actually runs on it today

The clearest example is video calling. Microsoft's support page for Windows Studio Effects says the feature requires an NPU and, for the full set of effects, a machine in the 40 TOPS tier. Blur, framing and noise suppression run continuously for the whole call, which is exactly the kind of steady, modest workload an NPU handles for a fraction of the power a GPU would use. Live captions, some photo editing effects and screen-content search follow the same pattern. On Macs, Apple's privacy page for its built-in AI features says some models run entirely on the device, while more complex requests are sent to Apple's own server-based models.

What still needs the GPU or the cloud

Large language models are limited by how fast the laptop can read memory, not by raw operations, so a big NPU number does little for a local writing assistant. The guide to how much memory a local model really needs covers that sizing. Heavier image generation and video work still favour a capable GPU. And many features branded as AI, including chat assistants, web summaries and writing help inside online apps, run on remote servers whatever chip your laptop has. The NPU does not change what those services receive; their privacy settings do.

How to check what a laptop has

• Windows: open Task Manager and the Performance tab. Laptops with an NPU list it alongside CPU, memory and GPU, with its own usage graph.
• Spec sheets: look for the NPU TOPS figure separately from the combined platform TOPS, which adds CPU and GPU numbers together and looks much larger.
• Features: the manufacturer's page usually lists which on-device features are supported. Features are what you use; TOPS is not.

To weigh the NPU against everything else on the page, it helps to read a laptop spec sheet line by line. If writing is your main job, the advice on choosing a laptop for writing puts the NPU in its place. The NPU also powers local speech features, which matters if you care about transcribing recordings without an internet connection, and some of the features it runs are ones you may prefer off, which is covered in the guide to turning off the built-in AI features. For every other buying question, start from the laptop buying guide hub.

Common questions

Do I need an NPU laptop?

Only if you want the specific on-device features that require one, such as call effects on Windows. For writing, browsing and office work, a laptop without one is fine.

Is a higher TOPS number always better?

No. It is a peak figure at low precision. Memory, software support and cooling decide real performance, and many apps do not use the NPU at all.

Will an NPU make local AI writing models faster?

Rarely today. Those models are limited by memory bandwidth and usually run on the GPU or CPU.

Does an NPU keep my data private?

It lets some features run on the device. Features that call cloud services still send data out, whatever hardware you have.

Does the NPU drain the battery?

It is designed to do the opposite: run steady AI tasks with less power than the CPU or GPU would need.