NPU

in English: neural processing unit

basicwidespread

The NPU is the part of the phone or laptop chip specifically designed for artificial intelligence calculations, performing them locally, quickly, and with low power consumption.

In plain terms

Your computer's processor (CPU) is like a skilled handyman: it can do a bit of everything, but only a few things at a time. Artificial intelligence, however, needs a huge number of small, identical multiplications and additions done at once. For that job the handyman is slow and uses a lot of power.

By contrast, the NPU is like a kitchen robot: designed for a single type of task, which it performs quickly and efficiently. Such tasks include blurring the background in video calls, speech recognition, enhancing photos, or running a small language model.

The most important benefit is that the AI runs on your own device, not on a remote server. This means it works without internet, has no network latency, and if the function truly runs entirely locally, your data never leaves the machine. Compared to the processor or graphics unit, the NPU performs the same work with far less power, so the battery lasts longer.

The NPU is usually not a separate component; it is mostly built into the same chip that contains the processor and graphics unit.

How it works

An AI model is essentially a huge set of tables of numbers; running it means multiplying and adding these tables together many, many times.
In the NPU, many small computing units work in parallel, specifically designed for these operations, so they perform simultaneously what the processor would largely do sequentially.
The model is usually 'compressed' beforehand: high-precision numbers are rounded to smaller, simpler numbers (this is quantization) because the NPU works fastest and with least power using these.
The operating system and the application decide where each task goes: general work to the CPU, demanding graphics and often larger local AI models to the GPU, and continuously running, smaller AI tasks to the NPU.
Performance is often given in TOPS, which shows how many trillions of operations are performed per second; this is the number you see on the box in the store.

An everyday example

During a video call, you enable background blur, yet the laptop's fan does not spin up and the battery does not drain within an hour. The NPU performs this work quietly while the processor remains free for your other programs.

Who is affected

everyoneWhen buying a new phone or laptop, the presence and power of the NPU, along with memory size, largely determine which AI functions will run locally on the device without internet.

employeeOn corporate devices, the NPU enables certain note-taking, translation, or summarization functions to run locally, preventing confidential material from being sent to the cloud. However, many such functions still operate in the cloud even on NPU-equipped devices, so this should be verified. → Work & HR

entrepreneur, executiveWhen buying a computer, consider that AI running locally can reduce cloud subscription costs and data privacy risk.

developerIf your application or agent runs a model locally, you must optimize it for the NPU and account for manufacturer-specific toolkits. → Code

A common misconception

MythIf my device has an NPU, large chatbots run locally on my machine, making all AI faster.

In realityThe NPU is smaller and intended for local tasks; the largest models still run in data centers, and the NPU only accelerates the program specifically prepared for it.

Limits and risks

Not enough for the largest language models: they need far more memory and computing power than fits in a phone or laptop.
It is only valuable if the software uses it; many programs do not touch the NPU at all, so the number on the box is not a guarantee on its own.
TOPS values are calculated using different methods by manufacturers, so two machines cannot be reliably compared based solely on this.
The NPU is for running pre-trained models, not for training new ones.

Similar, but not the same

GPU (graphics processing unit) – The GPU is more versatile and powerful, suitable for training models, but consumes significantly more power; the NPU is designed for narrower tasks, with energy efficiency as its main advantage.
CPU (processzor) – The CPU can run any program, but it is slow and wasteful in AI's massive, uniform calculations, while the NPU is specialized for these and performs them very well.
AI accelerator – AI accelerator is the broader collective term for all hardware tailored to AI, including data center chips; NPU typically refers to the version built into consumer devices.

Related concepts

Broader

AI accelerator, hardver, processzor

Narrower

–

On the edge

GPU, on-device AI, edge computing, inference, quantization

A brief history

The NPU reached the mass market in 2017, when Huawei built a dedicated neural unit into its Kirin 970 chip and Apple into its A11 Bionic. For personal computers, it became a buzzword in 2024, when Microsoft required an NPU of at least 40 TOPS for the Copilot+ PC category.

Where things stand

At the end of 2026, nearly all new mid-range and high-end smartphones, and most new laptops, have an NPU. Hardware has outpaced software: there are still few everyday applications that truly make use of it.

Status: 2026 November

In the news

Appears in 1 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Critical-mineral and AI-chip supply alarms point to deeper supply-chain risksOct 19, 2026 · Ipar

Also known as

NPU · neural processing unit · neural processing unit · neural processor · AI-chip · AI accelerator · AI accelerator · Neural Engine

Sources

Wikipedia: Neural processing unit

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.