Every new laptop in 2026 advertises an NPU and a big TOPS number, and many are sold as “AI PCs” or “Copilot+ PCs”. But what is an NPU, what does it actually do, and does it help you run ChatGPT-style models or image generators on your own machine? This guide explains NPUs in plain English, compares the 2026 chips from Qualcomm, Intel, AMD, Apple and NVIDIA, and tells you when an NPU matters and when the GPU and memory matter far more.
- NPU = Neural Processing Unit, a low-power chip block built for AI maths. It runs small AI features (camera effects, captions, on-device assistants) without draining the battery.
- TOPS = trillions of operations per second, usually at INT8. It is a peak number, not a speed guarantee.
- Copilot+ PC = Windows laptop with a 40+ TOPS NPU, at least 16 GB RAM and 256 GB storage.
- 2026 NPUs: Snapdragon X2 Elite 80 TOPS, Ryzen AI 400 60 TOPS, Ryzen AI Max (Gorgon Halo) 50-55 TOPS, Intel Panther Lake about 50 TOPS.
- For local LLMs and image generation, RAM size, memory bandwidth and the GPU matter far more than NPU TOPS.
What Is an NPU?
An NPU (Neural Processing Unit) is a part of a processor designed to do one thing very efficiently: the matrix multiplications that neural networks are made of. CPUs are flexible but do this slowly. GPUs do it fast but use a lot of power. An NPU sits in between: much faster than the CPU for AI work, and much more power-efficient than the GPU, which makes it ideal for AI features that run all the time on a laptop, like background blur on video calls or live captions.
Apple has called its NPU the Neural Engine since 2017. Qualcomm’s is the Hexagon NPU, Intel’s is simply the NPU (formerly VPU), and AMD’s is built on its XDNA architecture. They all do the same job.
What Does TOPS Mean?
TOPS means trillions (tera) of operations per second. NPU makers usually quote it at INT8 precision, at peak clock, with every unit busy. It is a useful way to compare NPUs from the same generation, but it says nothing about memory bandwidth, software support or how a real model performs. Two warnings:
- NPU TOPS vs platform TOPS. Some brands add CPU + GPU + NPU TOPS together for a bigger headline number. Copilot+ certification only counts the NPU.
- GPU numbers are often FP4 or FP8. NVIDIA rates its RTX Spark platform at up to 1 petaflop of FP4, which is not directly comparable to an INT8 NPU figure.
NPU vs GPU vs CPU
| CPU | NPU | GPU | |
|---|---|---|---|
| AI speed | Slow | Medium | Fast to very fast |
| Power use | Medium | Very low | High |
| Best for | Small models, general tasks | Always-on AI features, small models on battery | LLMs, image and video generation, training |
| Software support for local AI | Universal | Limited, vendor-specific | Excellent (especially NVIDIA CUDA) |
What Is a Copilot+ PC?
Copilot+ PC is Microsoft’s label for Windows laptops that meet a minimum AI spec: an NPU with at least 40 TOPS, at least 16 GB of RAM and at least 256 GB of storage. These PCs unlock Windows features that run on the NPU, such as Recall, Click to Do, Live Captions with translation, Windows Studio Effects for the camera, and improved Windows search. All the main 2026 laptop chips from Qualcomm, Intel and AMD meet the NPU requirement.
2026 AI PC Chips Compared
| Chip | NPU | Max memory | Notes |
|---|---|---|---|
| Qualcomm Snapdragon X2 Elite / Extreme / Plus | 80 TOPS (all SKUs) | up to 48 GB on Extreme demo units | Arm Windows; about 225 GB/s on Extreme, about 150 GB/s on others (per Qualcomm interview) |
| Intel Core Ultra Series 3 (Panther Lake) | about 50 TOPS (reported) | 32 GB in tested systems | x86; Arc B390 graphics on Core Ultra X parts; first systems 27 Jan 2026 |
| AMD Ryzen AI 400 | 60 TOPS | – | x86; CES 2026; desktop parts from 2 Mar 2026 |
| AMD Ryzen AI Max (Gorgon Halo) | 50-55 TOPS | up to 192 GB (160 GB as VRAM) | Big 40-CU iGPU; built for local AI workstations |
| Apple M5 family | 16-core Neural Engine | 32-512 GB (M5 to M5 Ultra) | Neural Accelerators in every GPU core; Apple quotes about 4x AI performance vs M4 |
| NVIDIA RTX Spark (N1X) | – | up to 128 GB unified | Arm Windows with a Blackwell RTX GPU, up to 1 PFLOP FP4; laptops on sale 16 Oct 2026 |
Sources: Shattered’s November 2026 comparison, Laptops.computer’s AI PC buyer’s guide (27 September 2026), Tom’s Hardware for Gorgon Halo and PromptQuorum for Apple M5. Intel’s NPU figure is reported but not confirmed on Intel’s own pages, so treat it as approximate.
What Actually Runs on the NPU?
- Windows AI features: Studio Effects (blur, eye contact, auto framing), Live Captions, Recall, Click to Do and on-device Copilot features.
- App features: noise suppression in calls, photo editing tools (subject select, denoise), video effects in apps that use Windows ML / ONNX Runtime.
- Small language models: Microsoft’s Phi Silica and some vendor-optimised small models run on the NPU through tools like Microsoft Foundry Local, Intel OpenVINO, AMD Ryzen AI Software and Qualcomm AI Hub.
What usually does not run on the NPU: Ollama, LM Studio and llama.cpp in their default setups (they use the GPU and CPU), ComfyUI, and almost all image and video generation tools.
Do NPUs Help Run Local LLMs?
A little, but less than the marketing suggests. Generating text from an LLM is limited by memory bandwidth, not compute: every token has to read the model’s weights from memory. An 80 TOPS NPU on a laptop with about 150 GB/s of memory bandwidth can only generate as fast as that memory allows, exactly like the integrated GPU sharing the same memory. NPUs help with prompt processing and power efficiency, and they are great for small models on battery, but they do not make a laptop run a 30B model like a desktop GPU.
For local LLMs, these specs matter far more:
- Memory size (32 GB minimum for useful models, 64-128 GB for big ones). See how much VRAM you need for AI.
- Memory bandwidth (Apple M5 Max 460-614 GB/s, Ryzen AI Max about 256 GB/s, typical laptop 100-150 GB/s).
- A capable GPU with good software support (NVIDIA RTX, Apple Silicon, or AMD’s big iGPUs).
If local AI is your main reason for buying, compare Mac vs PC for local AI and the best GPUs for local LLMs.
How to Check If Your PC Has an NPU (and Whether It Is Being Used)
- Press Ctrl + Shift + Esc to open Task Manager.
- Click Performance. On recent Windows 11 versions with supported drivers, an NPU entry appears below the CPU, memory, disk and GPU graphs.
- Click it to see live NPU utilisation. Turn on a camera effect in a video call or Live Captions and you should see it rise.
- No NPU entry? Update the chipset and NPU drivers from your laptop maker (Intel, AMD or Qualcomm driver pages), then check again. Older processors simply do not have one.
You can also open Device Manager and look for a Neural processors category. On a Mac, every Apple Silicon chip has a Neural Engine, and Activity Monitor does not show it separately.
Try It: Run a Small Model on Your NPU
If you have a Copilot+ PC and want to see the NPU do real work, Microsoft’s Foundry Local is the easiest route. It picks the best version of a model for your hardware (NPU, GPU or CPU) automatically.
winget install Microsoft.FoundryLocal
foundry model list
foundry model run phi-4-miniThe first run downloads the model; after that you can chat in the terminal. Watch the NPU graph in Task Manager while it answers. Model names change as Microsoft adds new ones, so pick from foundry model list. Intel users can also try OpenVINO, AMD users the Ryzen AI Software, and Snapdragon users Qualcomm AI Hub models.
Be realistic: these are small models (a few billion parameters). They are great for summarising, rewriting and quick questions offline, but they are not a replacement for a 30B model on a desktop GPU.
Which AI Laptop for Which Person?
| You mainly want… | Look for | Memory | Skip |
|---|---|---|---|
| Office work, calls, battery life, Windows AI features | Any Copilot+ laptop (Snapdragon X2, Panther Lake, Ryzen AI 400) | 16 GB is fine, 32 GB better | Paying extra for a discrete GPU |
| Students on a budget | Copilot+ laptop on sale, or last year’s model | 16 GB | Chasing the highest TOPS number |
| Local LLMs on the go | MacBook Pro M5 Pro or Max, or Ryzen AI Max | 64-128 GB unified | 16 GB models of any brand |
| ComfyUI images and video | NVIDIA RTX 50 laptop GPU | 32 GB RAM plus 12-16 GB VRAM | NPU-only machines |
| Developers testing on-device AI | A Copilot+ laptop with your target chip vendor | 32 GB | Buying before checking which vendor SDK your app needs |
Buying an AI Laptop in 2026
- For everyday AI features and battery life: any Copilot+ laptop. Snapdragon X2 leads on NPU TOPS and battery; Intel Panther Lake and AMD Ryzen AI 400 give better x86 app compatibility.
- For running local LLMs on the go: a MacBook Pro with M5 Pro or M5 Max and as much memory as you can afford, an RTX 50 laptop, or a Ryzen AI Max laptop with 64-128 GB.
- For image and video generation: a laptop with an NVIDIA RTX GPU and at least 12-16 GB of VRAM.
- Memory is expensive this year. Laptop makers have raised prices and cut base memory because of the DRAM shortage. Do not buy 16 GB if you plan to run local models; see why RAM is so expensive in 2026.
Frequently Asked Questions
What is an NPU in a laptop?
An NPU (Neural Processing Unit) is a low-power processor block built for AI maths. It runs AI features like camera effects, live captions and small on-device models efficiently, so they do not drain the battery.
How many TOPS do I need for an AI PC?
Microsoft’s Copilot+ PC standard requires at least 40 NPU TOPS. All the main 2026 laptop chips exceed that, from about 50 TOPS (Intel Panther Lake) to 80 TOPS (Snapdragon X2 Elite).
Is an NPU better than a GPU for AI?
It is more power-efficient, but a GPU is much faster and has far better software support. For LLMs, image and video generation, the GPU and memory matter more.
Can I run Ollama on an NPU?
Not in its default setup. Ollama uses the GPU and CPU. To use an NPU you need vendor tools such as Microsoft Foundry Local, Intel OpenVINO, AMD Ryzen AI Software or Qualcomm AI Hub, usually with specially optimised models.
Is a Copilot+ PC worth it?
If you want the Windows AI features and long battery life, yes. If your goal is running big local models or generating images, spend your money on memory and a GPU instead.
How do I see NPU usage in Windows 11?
Open Task Manager > Performance. If your processor and drivers support it, an NPU graph appears there.
Does my phone have an NPU?
Almost certainly. Recent Snapdragon, Google Tensor, Apple A-series and MediaTek Dimensity chips all include one, and it powers features such as live translation, photo editing and on-device assistants.
Is 40 TOPS enough?
For Copilot+ features, yes, that is the requirement. Extra TOPS mainly helps future on-device features and developers; it does not change what you can do today very much.
Final Thoughts
NPUs are real and useful: they make everyday AI features fast and battery-friendly. But TOPS is the new megahertz, a big number that hides what matters. For serious local AI, memory and the GPU decide what you can run. If you already have a machine, start with installing Ollama and see what your hardware can do.
Sources
- Shattered: Snapdragon X2 Elite vs RTX Spark vs Panther Lake vs Ryzen AI 400 (3 Nov 2026)
- Laptops.computer: AI PC buyer’s guide 2026 (27 Sep 2026)
- Tom’s Hardware: AMD Ryzen AI Max 400 Gorgon Halo
- PromptQuorum: Apple Silicon M5 for local LLMs
- Microsoft Learn: Copilot+ PCs developer guide (NPU devices)
