What an NPU Actually Does: A Plain-English Guide to AI PC Chips
NPUs are now standard in new laptops. Here's what they actually do, how they differ from a CPU or GPU, and whether you should care.
In this article 8 sections
An NPU (neural processing unit) is a dedicated chip built to run AI workloads — like image recognition or local language models — using far less power than a CPU or GPU doing the same job. Most laptops sold since 2024 include one, but most owners never touch a setting that actually uses it.
This guide walks through what an NPU does, how it differs from the CPU and GPU already in your machine, and how to tell whether the NPU in a laptop you're shopping for actually matters for how you'll use it. It's written for anyone buying a new laptop or desktop in the next year, not just engineers.
What Is an NPU, Exactly?
A neural processing unit, or NPU, is a chip designed specifically for the math behind AI models — mostly matrix multiplication, run over and over on small chunks of data. That's a different job than a CPU's, which is built to handle many different kinds of instructions one after another.
According to Wikipedia's entry on neural processing units, this class of chip has existed in smartphones since the mid-2010s, well before it showed up in laptops. Apple, Qualcomm, and Google all shipped phone NPUs years before Intel and AMD put one in a Windows laptop chip.
In practice, the marketing term "AI PC" mostly just means "this laptop has an NPU built into the main processor." The NPU itself doesn't replace the CPU or GPU — it sits alongside them, on the same piece of silicon, and takes over specific AI tasks when the operating system routes work to it.
How Is an NPU Different From a CPU or GPU?
Each of the three chip types is good at a different kind of work. A CPU handles general logic and sequencing. A GPU handles large parallel workloads like graphics or training AI models. An NPU handles inference — running an already-trained AI model — using the least power of the three.
| Component | Best At | Typical Power Draw | Common AI Use |
|---|---|---|---|
| CPU | General logic, sequencing tasks | 15–45W | Light, occasional AI calls |
| GPU | Parallel graphics and training | 30–150W | Heavy local AI generation |
| NPU | Sustained, low-power inference | 1–10W | Background AI features, always on |
That power gap is the entire reason NPUs exist in laptops. Running a background AI feature like live captioning or webcam framing on the GPU all day would drain a battery in a few hours. On Copilot+ laptops, NPU-offloaded tasks can cost less than a tenth of the GPU equivalent.
How Many TOPS Does an AI PC Actually Need?
Most mainstream laptop buyers need 40 TOPS or more only if they want the official Copilot+ PC label and its exclusive features. Below that threshold, the NPU still works fine for everyday background tasks like noise cancellation and webcam effects — you just lose access to a handful of Microsoft's newest, more demanding AI features.
TOPS stands for trillion operations per second, and it's the standard yardstick vendors use to advertise NPU speed. Per Wikipedia's page on Copilot+ PCs, Microsoft set 40 TOPS as the minimum NPU performance required to carry that branding, a bar first cleared by Qualcomm's Snapdragon X chips in 2024 and later matched by Intel and AMD.
A higher TOPS number mostly buys headroom for future software, not a dramatically better experience today. Real-world differences between a 40 TOPS chip and a 55 TOPS chip in the same laptop generation tend to be barely noticeable for most current features.
Battery life is where the gap shows up more clearly than raw responsiveness. In one comparison across three Copilot+ PC laptops, running the same video-call session with NPU-accelerated background blur enabled used roughly 15% less battery over two hours than routing the same effect through the GPU on an older chip. That difference compounds over a full workday, which is the more practical reason to care about TOPS than any single benchmark number.
What Actually Runs on the NPU Today?
In 2026, the NPU mostly handles smaller, always-on features rather than full chatbot-style AI. Here's what typically routes to it on a current Windows or Mac laptop:
- Live video call effects — background blur, eye contact correction, auto-framing
- Background noise removal during calls and recordings
- Real-time captioning and translation
- Local image search and photo organization
- Small on-device language model features, like Windows' local Copilot actions
Running a full large language model locally — something closer to a full chatbot — still mostly happens on the GPU, or in the cloud, because most laptop NPUs don't yet have enough memory bandwidth for it. That's shifting fast, and it's a space Protocol Wire tracks closely as chipmakers push more memory onto AI PC silicon each generation.
The ENERGY STAR computers program has also started factoring AI-accelerator efficiency into its newer certification criteria, which is one sign regulators expect NPU-offloaded features to become standard rather than a premium extra.
Do You Need an NPU Right Now?
If your laptop workload is mostly browsers, office apps, and video calls without heavy AI filters, skipping the NPU costs you almost nothing today — but it will likely matter more within two to three years as more software routes AI tasks to it by default.
Buy for the NPU specifically if you're on frequent video calls and want longer battery life during them, or if you plan to keep the laptop three-plus years and want it ready for AI features that haven't shipped yet. Skip worrying about it if you mostly need a laptop for browsing, office work, or coding without local AI tools.
Related reading from Protocol Wire's hardware desk: our look at how large-scale AI chip procurement actually works covers the datacenter side of this same shift, and our report on XPeng's humanoid robot production line shows where on-device inference chips are headed next.
Which Laptop Chips Have the Strongest NPUs Right Now?
Every major chipmaker now ships an NPU inside its mainstream laptop processor line, and the TOPS numbers have converged more than most buyers expect. Here's how the current generation stacks up on paper.
| Chip Family | Maker | NPU Performance | Copilot+ Eligible |
|---|---|---|---|
| Snapdragon X | Qualcomm | ~45 TOPS | Yes |
| Core Ultra 200V (Lunar Lake) | Intel | ~48 TOPS | Yes |
| Ryzen AI 300 | AMD | ~50 TOPS | Yes |
| M4 | Apple | ~38 TOPS | Not applicable (macOS) |
The gap between these chips matters less than the marketing suggests. Real-world background-task performance — the kind that actually shows up in daily use — tends to differ little across this generation once a chip clears the 40 TOPS Copilot+ threshold (for the datacenter side, see the chips powering this year's AI boom). Where the difference actually shows up is battery life during sustained AI use, not raw feature availability.
Apple's approach is worth calling out separately: macOS doesn't use the Copilot+ branding at all, but Apple's Neural Engine has shipped in every Mac chip since the M1 and now handles a similar set of background AI tasks — on-device photo search, dictation, and Apple Intelligence features introduced in recent macOS releases.
A Short Buying Checklist for the NPU Specifically
If the NPU is actually a factor in your purchase decision, a few questions cut through most of the marketing noise faster than comparing raw TOPS numbers alone.
- Does the laptop carry the Copilot+ PC label, or is that irrelevant to you (macOS, Linux, older Windows workflow)?
- Do you spend multiple hours a day on video calls where background AI effects run continuously?
- Do you plan to keep this laptop long enough (three-plus years) for NPU-dependent software to mature further?
- Does your actual workload include any local AI tools today, or is it a hypothetical future use?
Answering "no" to most of these is a completely reasonable reason to deprioritize the NPU spec entirely and put that budget toward more RAM, a better display, or a larger battery instead — all of which affect daily use more directly for most buyers.
Conclusion
An NPU is a low-power AI chip that now ships inside most new laptops, handling background AI tasks the CPU and GPU would otherwise burn far more battery running. For most buyers in 2026, it's a nice-to-have rather than a must-have — but that balance is shifting as more everyday software starts routing work to it by default. Protocol Wire will keep tracking TOPS numbers and real-world battery results as the next generation of AI PC chips ships.
Protocol Wire covers the tools shaping how software actually gets built and shipped, for engineers and IT buyers who need the practical read, not the hype — with a bias toward the specs that affect daily use.
