A neural processing unit (NPU) is a specialized processor designed to run artificial intelligence tasks quickly while using very little power. It sits alongside the CPU and GPU in a computer and takes over AI workloads, such as background blur, voice isolation, and live captions, so the other processors are free for everything else.
NPUs are the defining component of an AI PC. Because they are built for one type of math, they can process AI features locally on your device, often with better battery life than running the same task on a CPU or GPU.
This guide explains how an NPU works, what TOPS ratings mean, and how an NPU compares with a CPU and GPU, so you can judge what matters when you choose a laptop.
How does an NPU work?
An NPU works by performing huge numbers of simple math operations at the same time, which is exactly what AI models need. Most AI tasks come down to multiplying and adding large grids of numbers, called matrices, over and over. An NPU contains many small compute units arranged to handle this matrix math in parallel.
Three design choices make NPUs efficient for AI:
Parallel matrix engines. Dedicated multiply-accumulate units process many calculations in a single step instead of one after another.
Lower-precision math. NPUs often use smaller number formats, such as 8-bit integers (INT8), which take less energy and memory than the higher-precision formats CPUs typically use. For most AI features, the accuracy difference is negligible.
On-chip memory. Keeping data close to the compute units reduces how often the NPU has to fetch information from system memory, which saves both time and power.
NPUs focus on inference, which is running a trained AI model to produce a result. Training large models from scratch still happens mostly on powerful GPUs in data centers. Learn more about training and inference in our machine learning guide.
What does TOPS mean?
TOPS stands for trillions of operations per second, and it measures the peak number of AI calculations a processor can perform each second. An NPU rated at 45 TOPS can, in theory, complete 45 trillion operations every second. A higher TOPS number generally means the NPU can run larger AI models or more AI features at once.
TOPS has become the headline spec for AI PCs because Microsoft uses it to define its Copilot+ PC category. A Copilot+ PC needs an NPU that can perform more than 40 TOPS, and many of the newest Windows AI features require that level of NPU performance (Microsoft Learn).
Typical NPU TOPS ratings by generation
Laptop generation | Typical NPU rating | Examples |
|---|---|---|
2026 AI PCs | 45–55 TOPS, some Arm-based chips up to 80 TOPS | Latest Copilot+ PC processors |
First Copilot+ PCs (2024–2025) | 40–50 TOPS | Qualcomm Snapdragon X (45), Intel Core Ultra 200V (48), AMD Ryzen AI 300 (up to 50) |
Early AI laptops (2023–2024) | About 10–16 TOPS | Intel Core Ultra "Meteor Lake" (10), AMD Ryzen 8040 (16) |
Sources: PC-WELT, Tom's Hardware.
How to read a TOPS rating
TOPS is a useful starting point, but it does not tell the whole story. Keep these points in mind when you compare laptops:
It is a peak number. TOPS describes maximum theoretical throughput. Real-world speed also depends on memory bandwidth, software optimization, and power and cooling limits.
Precision matters. TOPS figures are usually measured with 8-bit integer (INT8) math (Tom's Hardware). A figure based on lower-precision math, such as INT4, can look higher without the chip being faster at the same task.
NPU TOPS and platform TOPS differ. Some spec sheets list combined TOPS for the CPU, GPU, and NPU together. For Copilot+ features, the number that counts is the NPU's own rating.
Software support decides what you get. An NPU only helps when apps are built to use it. Check which features a laptop supports, not just its TOPS figure.
CPU vs. GPU vs. NPU: what is the difference?
The CPU, GPU, and NPU each handle a different kind of work, and an AI PC uses all three together. The CPU is the general-purpose manager, the GPU is the high-throughput workhorse, and the NPU is the power-efficient AI specialist.
CPU | GPU | NPU | |
|---|---|---|---|
Full name | Central processing unit | Graphics processing unit | Neural processing unit |
Designed for | General tasks and running the operating system | Graphics and large-scale parallel math | AI inference |
How it works | A few powerful cores that handle varied tasks in sequence | Thousands of smaller cores working in parallel | Dedicated engines optimized for matrix math |
Strength in AI | Small, quick AI tasks and coordinating work | Heavy AI work, such as generating images or running large models | Sustained AI features with low power draw |
Power use for AI | Moderate to high | High | Low |
Everyday examples | Opening apps, browsing, spreadsheets | Gaming, video editing, 3D rendering, local image generation | Background blur, noise removal, live captions, on-device assistants |
When each processor takes the lead
CPU: Best for tasks that need quick responses and frequent switching, such as running your apps and handling light AI work that finishes in a moment.
GPU: Best for bursts of intense AI or graphics work. A dedicated laptop GPU can post far higher TOPS than an NPU, but it uses much more power to do it (Tom's Hardware).
NPU: Best for AI features that run continuously or for long stretches, such as video-call effects, where saving battery matters more than peak speed.
The operating system and apps decide which processor handles each task. In many cases, a single AI feature splits its work across more than one of them.
What does an NPU do on a laptop?
On a laptop, the NPU runs the AI features you use every day without draining the battery or slowing down your other apps. Common NPU-powered features include:
Video-call effects. Background blur, auto-framing that keeps you centered, and eye-contact correction run continuously during a call, which makes them ideal NPU tasks.
Audio cleanup. Voice focus and noise suppression filter out keyboard clicks, traffic, and background chatter in real time.
Live captions and translation. Speech can be transcribed and translated on the device as it is spoken.
Smarter search. Some systems let you search for files and settings by describing them in everyday language.
Image creation and editing. Tools that generate images, remove objects, or upscale photos can run locally on capable NPUs.
On-device AI assistants. Smaller language models can run on the NPU to summarize, rewrite, or answer questions without sending data to the cloud.
Many of these features are covered in more detail in our AI features guide. Because the NPU processes data locally, it also supports on-device AI, which can offer faster responses and keep more of your information on your device.
Do you need a laptop with an NPU?
You do not need an NPU to use cloud-based AI tools, such as web chatbots, but an NPU is worth having if you plan to keep your laptop for several years. As more Windows and app features move onto the device, an NPU helps you get those features with better battery life and less reliance on an internet connection.
An NPU is a strong fit if you:
Spend a lot of time on video calls and want effects that do not slow your system down
Work on battery often and want AI features that sip power
Prefer to keep sensitive data on your device instead of sending it to the cloud
Want access to Copilot+ PC features, which require an NPU with more than 40 TOPS
When you compare laptops, look at the NPU's own TOPS rating, the amount of memory (Copilot+ PCs require at least 16 GB of RAM, according to Tom's Hardware), and the specific AI features each model supports. Our AI PC guide walks through what else to look for.
NPU frequently asked questions
What does NPU stand for?
NPU stands for neural processing unit. The name comes from neural networks, the type of AI model these processors are designed to run. Some manufacturers use other names for the same idea, such as AI engine, AI accelerator, or neural engine.
How do I know if my laptop has an NPU?
On a Windows 11 PC, open Task Manager and select the Performance tab. If your computer has a supported NPU, it appears in the list alongside the CPU, GPU, and memory. You can also check your processor's specifications on the manufacturer's website.
Is a higher TOPS rating always better?
Not always. A higher TOPS rating means more peak AI capacity, but real-world performance also depends on memory, software support, and how well apps are optimized for the NPU. Above the 40 TOPS Copilot+ threshold, the features a laptop supports often matter more than a few extra TOPS.
Can I add an NPU to my current computer?
In most cases, no. Laptop NPUs are built into the processor chip itself, so you get one by choosing a laptop with a processor that includes it. External AI accelerators exist, but they are mainly aimed at developers and specialized systems.
Does an NPU improve battery life?
An NPU can extend battery life when you use AI features. It runs AI tasks using less power than a CPU or GPU, so features like video-call effects drain the battery more slowly. It does not reduce power use for tasks that do not involve AI.
Does an NPU help with gaming?
Not directly. Games rely mainly on the GPU and CPU. An NPU can help indirectly by handling background AI tasks, such as voice chat noise removal, so the GPU and CPU stay focused on the game.
What is the difference between an NPU and a TPU?
A TPU, or tensor processing unit, is an AI accelerator designed by Google, mainly for data centers that train and run large AI models. An NPU is a broader term for an AI processor and usually refers to the low-power AI engine built into laptops and phones.
Do smartphones have NPUs?
Yes. Smartphones have included NPUs for years to power features like photo enhancement, face unlock, and voice assistants. Laptop NPUs follow the same idea, scaled up for PC workloads.
Related AI glossary guides
AI glossary: Definitions of key AI terms in one place
AI PCs: What makes a computer an AI PC and what to look for
On-device AI vs. cloud AI: Where AI runs and why it matters
AI features: The everyday AI capabilities NPUs make possible
Machine learning: How AI models learn, including training and inference
AI privacy and security: How on-device processing affects your data
