What is an NPU?
An NPU is a neural processing unit: a part of a processor built to do the multiplication and addition that AI models are made of, using little power. Microsoft calls it a specialized computer chip for AI intensive processes like real time translations and image generation, and requires one that can perform more than 40 trillion operations per second (TOPS) for a Copilot+ PC.

Think of it as a third engine next to the CPU and the GPU. It sits inside the same chip, so you can't buy one separately for a laptop.
The NPU does one kind of job, and it does it while sipping power. That's the point. A video call that blurs your background all day shouldn't drain the battery or slow down your other work.
What is the difference between an NPU, a GPU and a CPU?
The CPU handles everything, the GPU handles huge amounts of parallel math, and the NPU handles one kind of AI math with the least power. That split is the one IBM, AWS and HP all describe, and it answers the searches for NPU vs GPU and NPU vs CPU.
| CPU | GPU | NPU | |
|---|---|---|---|
| Built for | Any task, one after another | Thousands of parallel calculations | The math inside neural networks |
| Strength | Flexibility | Raw throughput | Low power for steady AI work |
| Weakness | Slow at AI math | High power draw | Only helps software written for it |
| Typical AI job | Small or odd tasks | Training and heavy inference | Background blur, captions, noise removal |
IBM's NPU vs GPU explainer says NPUs are usually paired with GPUs, not used instead of them, with the NPU taking repetitive AI work so the GPU can do bigger jobs. Microsoft says the same about Windows: it decides which of the three engines gets each task. A Google TPU is a cousin built for data centers, covered in TPU vs GPU.
One correction to a popular page. IBM's explainer says GPUs "equipped with integrated NPUs" speed up editing. NPUs ship inside the main processor or system chip, not inside a GPU.
How many TOPS does each chip's NPU have?
Makers rate current NPUs from 38 TOPS (Apple M4) to up to 85 TOPS (Qualcomm Snapdragon X2 Elite), and all the Windows chips below clear Microsoft's 40 TOPS bar. The table uses only figures from each maker's own pages, as of November 2026.
| Chip | NPU | TOPS the maker states | Clears the 40 TOPS bar? |
|---|---|---|---|
| Apple M4 | 16 core Neural Engine | 38 | Not a Windows chip |
| Apple M5 | 16 core Neural Engine | Not stated | Not a Windows chip |
| Qualcomm Snapdragon X Elite | Hexagon NPU | Up to 45 | Yes |
| Qualcomm Snapdragon X2 Elite | Hexagon NPU | Up to 85, some chips 80 | Yes |
| Intel Core Ultra 200V | Fourth generation NPU | Up to 48 | Yes |
| AMD Ryzen AI 9 HX 370 | AMD XDNA engine | Up to 50 | Yes |
Two things the table hides. Apple's M5 spec page lists a 16 core Neural Engine and Neural Accelerators and gives no TOPS figure at all. Intel's launch page also quotes up to 120 total platform TOPS, which adds the CPU, GPU and NPU together. That's not an NPU number, so don't compare it to the others.
What runs on an NPU?
Mostly small, steady AI jobs: background blur and noise removal in calls, live captions, real time translation and image generation. HP lists the first three. Microsoft names translation and image generation.
The software has to be written for it. Microsoft says the NPU is a new hardware resource that needs software specifically programmed to use it, and that the Windows 11 Copilot+ features were designed for it. An app that doesn't target the NPU runs on the CPU or GPU instead.
Models also need a diet. Microsoft explains that many NPUs only support integer math in a lower bit format such as INT8, so an AI model must be converted, or quantized, before it runs there. That's why NPU features tend to use small models. A chat model with hundreds of billions of parameters doesn't fit.
Do phones have NPUs, and how is an NPU different from a TPU?
Yes, phones have them too. Wikipedia lists Apple silicon, Qualcomm, Samsung, Huawei and Google Tensor as smartphone processors that carry AI accelerators, and Apple's Neural Engine is its name for the same idea.
A TPU is Google's version for a different job. Wikipedia describes Google TPUs as accelerators in Google's cloud data centers, while consumer NPUs are meant to be compact and energy efficient and to run small models quickly, usually in low bit formats such as INT8, with speed quoted in TOPS. TPU vs GPU covers the data center side.
What is a Copilot+ PC and why does 40 TOPS matter?
A Copilot+ PC is a Windows 11 computer whose NPU can perform more than 40 TOPS, and the number is Microsoft's entry line for its on device AI features. Microsoft's developer guide says many new Windows AI features require an NPU that runs at 40+ TOPS.
The guide lists the Snapdragon X Elite machines first, then adds the AMD Ryzen AI 300 series and the Intel Core Ultra 200V series. Older AI laptops don't clear the line. HP puts the earlier Intel Meteor Lake NPU at about 10 TOPS and AMD's Ryzen 8040 at about 16.
The bar is a floor for Microsoft's own features. It doesn't say anything about how fast your favorite chatbot will be.
Do you need an NPU to use ChatGPT, Claude or Gemini?
No. ChatGPT, Claude and Gemini run their models in the vendors' data centers, and your device only sends the question and shows the reply. A laptop without an NPU gets the same answer as one with an NPU. Best AI app for Windows covers what a Windows machine needs for each assistant.
An NPU helps when a model runs on the device itself. That's the case for the Windows features above and for some small local models, which local LLM explains. For the models themselves, what an LLM is is the place to start.
What does TOPS mean and how do you read it?
TOPS counts trillions of operations per second, and it is a ceiling, not a speed test. HP calls it a peak throughput figure. AWS says to read it alongside the number format, because a chip rated at 50 TOPS on small integers is not doing 50 trillion operations of full precision math.
Three checks make a TOPS figure useful. First, confirm it is the NPU line, not total platform TOPS. Second, confirm the format, which is usually INT8. Third, check the wording on the maker's page: Intel and AMD both say up to, and the Qualcomm X2 Elite page notes that some chips are rated at 80 instead of 85.
AWS also lists memory bandwidth, software maturity and model portability as things to weigh. A smaller number on a chip with good software support can beat a bigger number on a chip your apps ignore.
Three real cases where the NPU number misleads
- A buyer compares total platform TOPS. Intel quotes up to 120 for the whole Core Ultra 200V platform and up to 48 for its NPU. A rival NPU rated 50 is not slower than a platform rated 120. Compare NPU to NPU.
- A developer ships an app and sees no NPU use. Microsoft's Task Manager shows NPU usage beside CPU and GPU, and its guidance says software must be written for the NPU. If the NPU graph stays flat, the app is running on the CPU or GPU.
- A model that is too big. Microsoft says NPUs usually need quantized INT8 models. A buyer who expects a 45 TOPS laptop to run a huge model locally learns that memory and model format decide the answer. HP notes that a higher TOPS number doesn't tell you the size of model a chip can run.
What questions should you ask an AI about NPUs?
These five prompts work in Claude, ChatGPT, Gemini or Whizi. Turn web search on, because chip figures change with every launch.
- "Compare the NPU in [my laptop model] with the NPU in [another model]. Use only each maker's own spec page, give the TOPS and the link, and say not stated where a page is silent."
- "My laptop has [chip name]. List three apps I already use that can run on its NPU, and say how you know each one does."
- "Explain TOPS to me in two sentences, then tell me why total platform TOPS and NPU TOPS are different numbers."
- "I want to run a [size] parameter model on a laptop with [chip] and [amount] of memory. Is the NPU or the GPU the better target, and what would I have to convert?"
- "Is an NPU worth paying extra for if I mostly use [list your apps]? Give me the strongest case for and against, with sources."
Where does Whizi fit?
Whizi is one app for asking Claude, GPT and Gemini the same question, with no NPU needed, because the models run in the cloud. Web search is on by default in each chat, so questions about this month's chips come back with sources.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro adds Claude, GPT and Gemini, side by side in Compare, and its first week is $1.99. A spec page is a better check than any chatbot, so open the maker's page for the final number.
What can an NPU not do?
It can't make an AI model smarter, and its TOPS rating is a peak figure, not the speed you will see. HP says TOPS is a peak throughput number and not a complete measure of real world speed. AWS lists memory bandwidth, software maturity and model portability as things to check beside TOPS.
It also can't help software that doesn't target it. NPUs are newer and often proprietary, which IBM gives as a reason for fewer community tools than GPUs have. Treat the number on the box as a starting point and read how the apps you use run on it.
- Read the NPU line on the maker's own page, not the total platform figure
- Check for 40 TOPS or more if you want Copilot+ features
- Open Task Manager and confirm an NPU entry exists
- Remember cloud chatbots run in data centers and don't need an NPU
- Compare memory and software support, not only TOPS
Frequently asked questions
What does TOPS stand for?
TOPS means trillions of operations per second. Microsoft's Copilot+ PC rule counts operations on the NPU, and a figure of 40 TOPS means more than 40 trillion per second at peak. HP says it is a peak throughput figure, not a full measure of speed in real use.
Is an NPU better than a GPU?
An NPU is better at steady, low power AI jobs and a GPU is better at heavy parallel work, so neither wins everywhere. IBM says NPUs are usually paired with GPUs, with the NPU taking repetitive AI tasks so the GPU is free for larger ones.
Do I need an NPU in my next laptop?
Not for ChatGPT, Claude or Gemini, which run in the cloud, as of November 2026. You need one for Copilot+ PC features, which Microsoft ties to an NPU above 40 TOPS, and it helps with local AI features such as background blur and live captions.
Does the Apple Neural Engine count as an NPU?
Yes. The Neural Engine is Apple's name for its NPU. Apple says the M4 version is capable of 38 trillion operations per second, and its M5 spec page lists 16 cores but gives no TOPS figure.