What the neural engine in your phone actually does all day

Phone spec sheets have quietly grown a third processor. There's the CPU that runs your apps, the GPU that draws things, and then a block with a name like neural engine that most people have never had explained to them. It isn't marketing. It's the reason a phone can watch you through the camera and understand what it's seeing without asking the internet for help.

It is a chip that is good at one kind of math

Neural networks mostly do one thing over and over: multiply a lot of numbers together and add up the results, in enormous batches. A CPU can do that, badly. A GPU can do it well but hungrily. A dedicated neural block does exactly that pattern and little else, which is precisely why it's efficient at it.

The practical consequence is battery. Running a model continuously — every frame of a live camera feed, for minutes at a time — is only reasonable if each inference is cheap in power. That's the whole trick. The hardware doesn't make on-device machine learning possible so much as it makes it possible to leave running.

What it makes possible inside an app

The clearest example we ship is Form Coach. You prop the phone on the floor, and it draws a skeleton over your body in the live preview, counts your reps, and scores the form of each one. That runs offline, with no account, on iPhone and Android. Nothing about your workout goes anywhere.

Every part of that experience depends on the inference being both fast and cheap. Fast, because feedback about a rep is useless after the rep. Cheap, because a set lasts minutes and the phone has to survive the session. The same hardware story sits behind on-device speech recognition and text recognition: features that used to imply a server and now don't.

The interesting thing about this hardware isn't that your phone can run a model. It's that it can run one continuously, for free, while in airplane mode.

What it doesn't do

It is not a small version of a chat assistant, and it doesn't make your phone capable of open-ended reasoning. Local models are narrow by design: pose estimation, transcription, recognition, classification. Ask a narrow model a general question and you get nonsense, quickly and efficiently.

It also doesn't come free to developers. A cloud feature can be fixed by renting a bigger model; an on-device feature has to actually work on the hardware in the user's hand, at the frame rate real usage demands. That constraint is the reason on-device features tend to be either genuinely good or obviously broken, with much less middle ground.

Why it matters to you, not just to engineers

Three things follow from a feature running on this chip instead of a server. It works with no signal. It costs nothing per use, so it doesn't need to be metered or rationed. And the data it processes — video of your body, audio of your voice, photos of your home — never leaves the device, because there's nowhere for it to go.

The gap between what a local model can do and what a hosted one can do keeps narrowing. Every year that continues, more features quietly move off the network and into your pocket, and the ones we build stay free, no subscriptions, ever. See what's running on-device today across our apps.