What Is an NPU? The AI Chip Your Kid's Phone Already Has

Your kid holds up their phone. The screen lights up and unlocks before they even touch it — face recognition, instant. They take a photo and the background blurs precisely around them — portrait mode, real-time. They point the camera at Spanish text on a restaurant menu and it translates into English on screen — live translation, no internet required.

All three of those things feel like magic. None of them are magic. All of them run on a chip inside the phone called the NPU — Neural Processing Unit.

It’s probably the most important chip your kids have never heard of. It’s in every flagship phone sold since 2020. It’s what allows AI to run on-device without draining the battery or requiring an internet connection. And the fact that almost no one talks about it — even in technology education — is a gap worth closing.

Why Nobody Talks About the NPU

The CPU gets discussed in every computer science class. The GPU became famous through gaming and then AI. But the NPU sits in a strange middle space: too new to be in textbooks, too specific to get mainstream press coverage, too invisible in its operation for most users to notice.

Here’s why it matters to understand: the NPU is evidence that AI has moved from the cloud to the device. When face unlock was first introduced on phones, it worked by sending your face to a server, running recognition remotely, and sending back a “yes” or “no.” That process took time and required connectivity.

Modern face unlock — using the NPU — runs entirely on the phone. No server involved. Instant response. Works in airplane mode. This shift from cloud AI to on-device AI is one of the most significant trends in consumer technology, and the NPU is the hardware that made it possible.

Explained Like You’re 5: The Specialist Who Knows One Thing Perfectly

Imagine a restaurant with three workers:

The general manager (CPU) handles everything complex: scheduling, resolving problems, managing the whole operation. They can do anything but are busy all the time.

The line cook (GPU) works in parallel — cooking many dishes at once, great at repetitive, high-throughput work.

The dessert specialist (NPU) does only one thing: makes the perfect dessert. They do it incredibly fast, use almost no energy doing it, and free up the general manager and line cook for other things.

The NPU is the dessert specialist. It does AI inference — running trained neural networks on input data — extremely efficiently. It does almost nothing else. But that one thing, done the right way, enables face unlock in under a second, real-time photo enhancement, and local language processing without ever needing to call out to the cloud.

How It Actually Works

A neural network is a mathematical model — a massive set of numbers (called weights) arranged in layers, trained to recognize patterns. When face unlock runs, it takes a camera image and runs it through hundreds of mathematical operations across those layers to determine: does this match the stored face? Yes or no?

These operations — matrix multiplications, activation functions — look like GPU workloads. But the NPU is different from the GPU in one crucial way: it’s optimized for inference (running trained models) rather than training (learning the weights in the first place).

Training a model requires floating-point math with high precision — that’s what GPUs do. Running a trained model for inference can use lower-precision math (INT8 or even INT4 — integers rather than floating-point numbers) and still get correct results. The NPU is built around this lower-precision, inference-focused math. It uses far less power than a GPU for the same inference task.

An Apple A17 Pro’s Neural Engine can perform 35 trillion operations per second (TOPS) while consuming a fraction of the power a GPU would need for the same task.¹ That’s the engineering: not raw power, but efficient power. A phone that could do face recognition using the GPU would drain its battery in hours. Using the NPU, it can do it thousands of times a day without measurable battery impact.

Modern NPUs also handle:

• Real-time photo processing (noise reduction, HDR fusion, portrait segmentation)
• On-device voice recognition (“Hey Siri,” “OK Google” — without cloud)
• Real-time translation using camera
• Autocorrect and predictive text (ML-powered)
• Health monitoring (heart rate, sleep analysis on smartwatches)

Why Kids Should Know This Today

On-device AI is not a trend — it’s an architectural shift. Apple, Google, Qualcomm, and Samsung all made explicit strategic decisions to build NPUs into their chips because they recognized that AI running locally (without cloud latency, without privacy concerns, without connectivity requirements) is fundamentally better for most everyday tasks.

A child who understands why on-device AI matters — the trade-offs between cloud and on-device, the role of the NPU, the difference between training and inference — understands the most important current development in consumer AI. Most adults who use these features daily cannot explain them.

The U.S. Bureau of Labor Statistics projects that computer hardware engineering roles will grow steadily through the 2030s, with semiconductor and AI hardware design driving the growth.² The companies building NPUs — Apple, Qualcomm, Google, MediaTek — are among the most strategically important technology companies in the world.

How to Teach Your Kid About This

Ages 5–8: The Specialist Helper Game

Play a sorting game. Give your child a pile of mixed-color blocks. Ask them to sort by color as fast as they can while also doing a math problem in their head. It’s hard — the two tasks compete.

Now assign a helper whose only job is sorting blocks. Your child does the math problem; the helper handles the sorting. The total work gets done faster, and neither is distracted by the other’s task.

That’s what the NPU does. It handles AI tasks so the CPU and GPU don’t have to, making everything run faster and more efficiently.

Ages 9–12: Find the NPU in Your Phone

Look up your phone’s chip and find out if it has an NPU. Here are the key chips to search:

• Apple: A-series chips (A14 and later) all have a “Neural Engine”
• Qualcomm: Snapdragon 8 series has a Hexagon NPU
• Google Pixel: Tensor chips have a dedicated Tensor Processing Unit (TPU)
• Samsung: Exynos chips have a dedicated NPU

Then test it: turn off Wi-Fi and cellular data. Open the camera and try portrait mode, face unlock, and real-time translation (in Google Lens or the camera translate feature). These all work offline. That means they’re running on the NPU, on-device — no cloud needed.

Ages 13+: Training vs. Inference — the Key Distinction

Help your teen understand the difference between training and inference — the most important technical distinction in practical AI.

Training: Feeding millions of examples into a neural network, adjusting the weights until the model gets good at a task. Takes enormous compute (GPUs, weeks of time, massive energy). Happens once (or periodically when the model is updated). Done in data centers.

Inference: Taking a trained model and running it on new input to get an output. Can be done with less compute. Happens millions of times a day on individual devices. This is what the NPU does.

Analogy: training is writing a cookbook. Inference is using the cookbook to cook one meal. You write the cookbook once; you cook from it every day.

For teens interested in AI: Google’s free Machine Learning Crash Course (ml.google.com) explains neural networks and inference with interactive examples. No prior experience needed.

NPU Implementations Across Major Phone Chips

Chip (Year)	Phone Examples	NPU Name	Performance	Key On-Device Features
Apple A18 Pro (2024)	iPhone 16 Pro/Max	Neural Engine (16-core)	35 TOPS	Siri on-device, photo AI, Face ID
Qualcomm Snapdragon 8 Elite (2024)	Samsung Galaxy S25, OnePlus	Hexagon NPU	45 TOPS	Real-time translation, photo processing
Google Tensor G4 (2024)	Pixel 9 series	TPU (Google custom)	~20 TOPS (est.)	Call Screening, live captions, photo unblur
Samsung Exynos 2500 (2025)	Galaxy S25 (select regions)	NPU (custom)	~30 TOPS	On-device Bixby, photo AI
MediaTek Dimensity 9400 (2024)	Various Android flagships	APU 890	35 TOPS	On-device AI, edge inference

TOPS = Trillions of Operations Per Second

The performance race in NPUs is moving faster than any other part of chip design. From the A11’s 600 billion operations per second in 2017 to today’s 35+ TOPS, that’s roughly a 50x improvement in seven years.³

This Chip in Devices Your Kid Uses

iPhone: The Neural Engine runs continuously in the background — Face ID, photo processing, Siri voice recognition, text predictions. It’s doing real AI computation dozens of times per minute.

Android flagships: Every Qualcomm Snapdragon 8-series phone has a Hexagon NPU. Every Google Pixel has a Tensor chip with dedicated ML hardware. On-device captioning, translation, and photo features all use it.

Smart speakers: Amazon Echo and Google Nest both use small NPU-like ML accelerators to recognize wake words (“Alexa,” “Hey Google”) locally, only sending audio to the cloud after the wake word is detected. This improves response speed and reduces the privacy footprint.

Apple Watch: The chips in recent Apple Watch models include ML accelerators specifically for real-time health analysis — heart rate irregularity detection, sleep staging, crash detection — all running on-device.

What to Watch for Over the Next 3 Months

Weeks 2–4: Your child should be able to explain one NPU-powered feature on their device and describe why it can run without an internet connection. “Face unlock works offline because the NPU processes the math locally” is a solid answer.

Month 2: Understanding the training vs. inference distinction is the next milestone. Can they explain why you can’t train a new model on your phone, but you can run an existing one? The computational and power constraints are the answer.

Month 3: A strong understanding shows when they can identify which features on a device are cloud-based AI (anything that requires connectivity) vs. on-device AI (features that work offline). This distinction is practically useful — it predicts latency, privacy exposure, and connectivity requirements all at once.

FAQ

Does my child’s phone have an NPU?

If it’s a flagship or upper-mid-range phone released after 2020, almost certainly yes. Apple A14 and later, Qualcomm Snapdragon 7 and 8 series, Google Tensor, Samsung Exynos 2100 and later, MediaTek Dimensity 9000 and later — all include dedicated NPU hardware. Budget phones may have limited or no dedicated NPU.

Is the NPU the same as the “AI chip” in marketing?

Yes. “AI chip,” “Neural Engine,” “Tensor chip,” “NPU,” and “ML accelerator” all refer to the same category of hardware. Marketing names vary; the underlying function is the same: running neural network inference efficiently on-device.

Can the NPU train AI models on my phone?

Not practically. On-device AI chips handle inference only. Training requires far more compute and memory than any current phone can provide. You can fine-tune very small models locally, but training a model from scratch happens in data centers on GPU clusters.

What does “on-device AI” mean for privacy?

It means your data doesn’t leave your phone for those features. Face unlock, on-device transcription, local translation — none of this sends your face, voice, or text to a server. That’s a meaningful privacy benefit. Features that do require cloud (like asking Siri a question that needs internet information) are separate from on-device AI inference.

Will NPUs replace GPUs?

No — they serve different roles. GPUs handle training and high-performance inference (like running ChatGPT on a server). NPUs handle efficient inference on constrained devices (phones, watches, earbuds). Both are growing in importance. The two aren’t competing so much as handling different parts of the AI workload.

How much battery does the NPU actually save?

Significant amounts for continuous tasks. Running face unlock using the GPU would consume orders of magnitude more power than the NPU. For intermittent tasks like taking a photo, the difference is smaller. But for always-on features — wake word detection, health monitoring, persistent translation — the NPU’s efficiency is what makes them practical to run continuously.

---

About the author Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Apple Inc. (2023). A17 Pro Chip Overview. https://www.apple.com/iphone-15-pro/a17-pro/
2. U.S. Bureau of Labor Statistics. (2022). Computer Hardware Engineers: Occupational Outlook Handbook. https://www.bls.gov/ooh/architecture-and-engineering/computer-hardware-engineers.htm
3. Apple Inc. (2017). A11 Bionic Chip — Neural Engine Overview. https://www.apple.com/newsroom/2017/09/iphone-x/
4. Qualcomm Technologies. (2024). Snapdragon 8 Elite Platform Technical Overview. https://www.qualcomm.com/products/mobile/snapdragon/smartphones/snapdragon-8-series-mobile-platforms/snapdragon-8-elite-mobile-platform
5. Dally, W. J., Turakhia, Y., & Han, S. (2020). “Domain-specific hardware accelerators.” Communications of the ACM, 63(7), 48–57. https://doi.org/10.1145/3361682
6. Google AI. (2024). Google Tensor G4 Chip Overview. https://store.google.com/us/category/phones

Footnotes

1. Apple A17 Pro Chip Overview, 2023. ↩

2. BLS Computer Hardware Engineers OOH, 2022. ↩

3. Apple A11 Bionic Overview, 2017. ↩