Short answer

  • The Neural Engine is Apple's dedicated accelerator for machine learning, built into its A-series and M-series chips. It runs trained neural networks (inference) far faster and more efficiently than the CPU or graphics processor could.
  • It first appeared in the A11 Bionic chip in 2017, with 2 cores performing up to 600 billion operations per second, powering Face ID and Animoji. By 2024 the M4 version reached 38 trillion operations per second, and the A20 Pro in 2026 doubled the design to two 16-core blocks.
  • It works alongside the CPU and graphics processor, not instead of them. Apple's Core ML software decides which of the three runs each part of a model.
  • Its point is on-device processing: answers without a round trip to a server, which means speed, offline use, lower energy per task, and, in Apple's framing, privacy, because personal data does not have to leave the device.

A neural engine is a piece of hardware built to do one job very fast: run neural networks, the mathematical models behind most modern artificial intelligence features. “Neural Engine,” capitalized, is Apple’s name for that hardware in its own chips. Since 2017 it has sat inside every A-series chip for the iPhone and iPad and every M-series chip for the Mac, quietly handling tasks such as recognizing a face, cleaning up a photo, or transcribing speech, while the main processor gets on with everything else.

The idea is not unique to Apple. Phone and computer makers now compete on this kind of silicon, usually called an NPU, short for neural processing unit. But Apple’s version is the one most people have heard of, because Apple put it in a phone in 2017 and named it. This article explains what the Neural Engine actually computes, how it grew from a two-core experiment in the iPhone X to the dual 16-core design in the iPhone 18 Pro, what it powers in daily use, and why the headline numbers attached to it need to be read with care.

What the Neural Engine Actually Computes

A neural network, stripped of the mystique, is arithmetic. A trained model is essentially a very large set of numbers, called weights, arranged in layers. Feed it an input, such as the pixels of a photo or the samples of a voice recording, and it produces an answer by multiplying the input values by the weights, adding the results together, passing them through simple rules, and repeating the process layer after layer until an output emerges: this is a face; this word was “yes”; this region of the picture is sky.

That means the dominant operation, performed billions of times for a single photograph, is a multiply-and-add step, known in the industry as a multiply-accumulate. General-purpose processors can do this arithmetic, but they are built for flexibility: following instructions, branching, managing memory, running code that was never anticipated. A neural engine is the opposite. It is a fixed array of arithmetic units laid out to stream these multiplications and additions through in massive parallel batches, using compact number formats that are precise enough for neural networks while being cheaper to move and process. It sacrifices flexibility for throughput per watt.

Two consequences follow. First, the Neural Engine is mostly an inference engine. Training a model, the process of adjusting the weights by showing the network millions of examples, is a much heavier job that Apple and other developers generally do on banks of server hardware before the model ever reaches a phone. The chip in the phone runs the finished model on new data. (Apple’s Core ML software has supported limited on-device training for certain kinds of models since 2019, so apps can personalize some models on the device, but that is the exception rather than the design center.) Second, the work is only useful if data can be fed fast enough, which is why the Neural Engine sits on the same piece of silicon as the CPU, the graphics processor and the memory controller, sharing the chip’s unified memory rather than waiting on copies of the data.

Where It Came From: Face ID and the A11

Apple introduced the Neural Engine in September 2017, inside the A11 Bionic chip that powered the iPhone 8, iPhone 8 Plus and iPhone X. In its launch release, Apple described a dual-core design performing up to 600 billion operations per second, built for specific machine learning algorithms. Its headline jobs were the two features that defined the iPhone X: Face ID, which matches a mathematical model of the owner’s face against the depth map captured by the front camera, and Animoji, which tracked more than 50 facial muscle movements to animate an emoji in real time.

The second generation showed the direction of travel. The A12 Bionic in 2018 expanded the design to eight cores and 5 trillion operations per second, and, significantly, Apple opened the Neural Engine to Core ML for the first time, so outside developers could run their own models on it rather than only Apple’s features using it. Core ML models on the A12 ran up to nine times faster than on the A11, Apple said, with as little as a tenth of the energy. From that point the Neural Engine stopped being a Face ID component and became a platform that every app could draw on.

The Generations, in Apple’s Own Numbers

Each year Apple publishes a peak figure for the new chip’s Neural Engine, measured in operations per second. The table below collects the figures Apple stated at each launch, with one exception, explained in the note below the table. Two patterns stand out. The core count has barely moved: the design jumped from 2 cores to 8, then to 16 in 2020, and stayed at 16 for six years while the speed rose through faster manufacturing processes and higher clocks. Then, in the A20 Pro of September 2026, Apple doubled the block itself, putting two 16-core Neural Engines on one chip.

ChipYearNeural Engine coresPeak operations per second
A11 Bionic20172600 billion
A12 Bionic201885 trillion
A13 Bionic201986 trillion
A14 Bionic20201611 trillion
A15 Bionic20211615.8 trillion
A16 Bionic202216Nearly 17 trillion
A17 Pro20231635 trillion
A18 / A18 Pro20241635 trillion
A19 Pro202516Not headlined by Apple
A20 Pro2026Dual 16-core (32 total)Not headlined by Apple

One figure in the table needs a caveat. For the A13, Apple published no peak operations number at launch: it headlined a 20 percent speed gain at 15 percent less power, and the “1 trillion operations per second” figure quoted at that launch described the CPU’s machine learning accelerators, not the Neural Engine. The 6 trillion shown for the A13 is the figure carried by standard reference tables for the chip, derived from the A12’s 5 trillion plus Apple’s stated 20 percent gain. It is a reasonable reference figure, but it is not a number Apple itself published for the Neural Engine.

The Mac line, which adopted the Neural Engine with the M1 in November 2020, tracked the same path: 11 trillion operations per second in the M1, 15.8 trillion in the M2 (2022), 18 trillion in the M3 (2023), and 38 trillion in the M4 (2024), which Apple described as 60 times faster than the first Neural Engine in the A11. The desktop Ultra chips double the block by joining two dies: the M1 Ultra’s 32-core Neural Engine was rated at 22 trillion operations per second, and the M2 Ultra’s at 31.6 trillion.

For the two most recent iPhone chips, Apple changed what it emphasizes. The A19 Pro release in 2025 described a 16-core Neural Engine working alongside new Neural Accelerators built into each graphics processor core, without a new peak figure. The A20 Pro release did the same for its dual design, saying the 32 total cores deliver double the AI processing power of the A19 Pro.

What It Powers in Daily Use

Most people meet the Neural Engine without knowing its name, through features that have to feel instant:

  • Face ID. Unlocking the phone means comparing a live depth map of a face against a stored mathematical representation, on the device, in a fraction of a second. This was the original job in 2017.
  • Photography. Apple’s computational photography, the processing that merges multiple exposures into one photograph, draws on the Neural Engine together with the image signal processor. Apple has credited the engine with a role in Smart HDR and in Deep Fusion, its texture-improving process, and with accelerating the machine learning behind features such as Portrait mode and, in the M4 launch material, Visual Look Up, which identifies objects in photos and video, and Live Captions, which transcribe audio in real time.
  • Siri and dictation. Speech recognition that once required a server round trip increasingly runs on the device itself, which is why dictation and many Siri requests now work quickly and can work offline.
  • Apple Intelligence. Apple’s generative AI features, introduced in 2024, are the heaviest load the Neural Engine has been given: summarizing text, rewriting drafts, generating images and answering questions about what is on screen. Apple publishes the hardware floor for these features, and it is expressed in Neural Engine generations: they require an iPhone with the A17 Pro or later, or an iPad or Mac with an M1 or later.
  • Third-party apps. Through Core ML, any developer can ship a model that runs on the Neural Engine: photo editors that select a subject in one tap, language apps that grade pronunciation, scanners that lift text out of images, fitness apps that classify movement.

Three Tools on One Chip: CPU, GPU and Neural Engine

An Apple chip carries three kinds of compute, and the division of labor explains why the Neural Engine exists at all.

The CPU (central processing unit) is the generalist. It runs the operating system and app logic, handles decisions and branching, and is fastest when a task is sequential and unpredictable. Apple also builds machine learning accelerators into its CPU cores (matrix-multiplication units known as AMX blocks, present since the A13), so the CPU is a capable fallback for model work.

The GPU (graphics processing unit) is built for thousands of parallel operations on pixels, and because neural network mathematics is also parallel, GPUs were the standard way to run models before dedicated accelerators spread. A GPU is more flexible than a Neural Engine and remains the better tool for some model shapes and for the largest models, but it draws more power for the same inference work.

The Neural Engine is the specialist: narrow, fixed in purpose, and the most efficient of the three at the multiply-and-add streaming that inference consists of. Software rarely chooses among them by hand. Apple’s Core ML framework examines a model and assigns its layers across the CPU, GPU and Neural Engine, and the chip’s scheduler can shift work between them; Apple’s A12 release described exactly this as a “smart compute system” deciding where each algorithm should run.

Why On-Device Processing Matters

Running a model on the device instead of in a data center changes four things.

Speed. There is no network round trip. Face ID, live transcription and photo effects respond in milliseconds because the computation happens where the data already is.

Availability. On-device features work in airplane mode, in a tunnel, or anywhere without coverage. The model travels with the phone.

Energy. Moving data costs power, and a radio sending photos to a server spends more of it than an accelerator spending a fraction of a second on inference. Efficiency was Apple’s stated reason for building dedicated silicon from the start: Johny Srouji, Apple’s hardware chief, explained at the A11 launch that custom silicon for neural processing performs the same tasks much more energy efficiently than a graphics engine.

Privacy, as Apple frames it. Apple’s consistent argument is that processing on the device keeps personal data on the device: Face ID data, voice recordings and photo analysis do not have to be sent to a server at all. For Apple Intelligence, Apple says many of its models run entirely on the device, and that requests needing larger models go to Private Cloud Compute, its server system designed so that, in Apple’s words, a user’s data is never stored or shared with Apple and is used only to fulfill the request, with the server code open to inspection by independent experts. Those are Apple’s claims about its own systems, and the company presents them as the reason for doing AI work locally wherever the hardware allows.

Not Just Apple: The Wider NPU Market

Apple’s branding is its own, but the hardware category is industry-wide, and the generic term is NPU. Qualcomm’s Snapdragon chips, used in most high-end Android phones, include a Hexagon processor with a fused AI accelerator that plays the same role. Google’s Tensor chips for its Pixel phones take their name from Google’s Tensor Processing Units (TPUs), the AI accelerators in its data centers, and carry an on-chip descendant of that technology for running Google’s models. The details differ, but the reasoning is identical everywhere: inference is too frequent, too latency-sensitive and too power-hungry to leave to a general-purpose processor.

Reading the Trillion-Operations Numbers Honestly

The peak figures in the table are real, but they are marketing-shaped numbers, and three cautions apply.

First, they are peaks: the theoretical maximum the hardware can sustain on its best-suited workload, not the speed of any particular app. Second, manufacturers do not all count an “operation” the same way, and the arithmetic precision behind a figure (whether each operation is performed on large or small numbers) can change the total severalfold, which is why TOPS figures cannot be compared across brands. Third, a doubled specification does not mean doubled real-world speed. When the A17 Pro doubled the A16’s headline figure to 35 trillion operations per second, independent benchmark testing by Macworld found genuine gains that were nonetheless well short of twice as fast, a reminder that memory bandwidth, model design and software often matter more than the accelerator’s ceiling.

The more meaningful question is usually not how fast the newest Neural Engine is, but which one a given feature requires. Apple’s own feature lines, such as the A17 Pro minimum for Apple Intelligence, tell users more than the trillion-operations race does.

Frequently Asked Questions

Does my iPhone or Mac have a Neural Engine?

If it has an A11 Bionic chip or later (iPhone 8 and iPhone X onward, and iPads and Macs from the same era forward) or any M-series chip, yes. Every A-series and M-series chip since 2017 includes one. Intel-based Macs and devices older than 2017 do not have one.

Is the Neural Engine the same as an NPU?

In substance, yes. NPU (neural processing unit) is the industry’s generic name for a dedicated machine learning accelerator. Neural Engine is Apple’s brand name for its NPU. Rivals use other names, such as Qualcomm’s Hexagon processor and Google’s edge TPU in its Tensor chips.

Does the Neural Engine train AI models?

Almost entirely, no. Training is done beforehand on server hardware, and the Neural Engine’s job is inference: applying the trained model to new photos, speech and text on the device. Apple’s Core ML has supported limited on-device training for certain model types since 2019, used mainly for personalization.

Why does Apple Intelligence need an A17 Pro or M1 at minimum?

Those are the floors Apple published when it introduced the features in 2024. Apple has not separated the requirement into a single specification; it reflects the combined performance of the chip, including its Neural Engine, graphics processor and memory. Older chips still run many machine learning features, but not the Apple Intelligence set.

Is the Neural Engine used for games?

Not directly. Games render on the graphics processor. The Neural Engine may support a game’s machine learning features, such as smarter non-player characters or voice processing, but frame rates are a GPU matter.

Conclusion

A neural engine is best understood as a concession to arithmetic. Modern AI features are, at bottom, billions of multiplications and additions, and doing that work on hardware shaped for exactly that pattern is faster and cheaper in energy than bending a general-purpose processor to the task. Apple’s Neural Engine is the most established example: introduced almost as an accessory to Face ID in 2017, it became the shared foundation for computational photography, on-device speech and, from 2024, generative AI features across the iPhone, iPad and Mac. The real story is the one the A20 Pro’s doubled block tells: the part of the chip reserved for AI is now a headline component, designed for as deliberately as the CPU itself.

Keep reading: What Is a Neural Network? How It Works, in Plain English

Sources & references

Emma Hudson

Editor of Amberaisle — clear, thoughtful explainers for curious readers. More about the editor →