What Is a Neural Engine? Apple’s NPU Explained in Detail

You’ve probably noticed your iPhone handles Face ID, Siri suggestions, and photo searches instantly, without slowing down. That speed isn’t magic — it comes from a piece of hardware called the Neural Engine, Apple’s name for a dedicated chip that accelerates machine learning tasks. This guide explains what the Neural Engine is, how it works across iPhone, iPad, and Mac, and why it matters for privacy and performance in the age of Apple Intelligence.

First introduced: Apple A11 Bionic (2017) ·
Current core count (latest): 16 cores (Apple M-series) ·
Primary function: accelerate machine learning tasks on-device ·
Devices equipped: iPhone, iPad, Mac (Apple Silicon) ·
Performance (16-core): up to 11 trillion operations per second

Six facts show one pattern: the Neural Engine evolved from a smartphone secret weapon into the brain of every Apple Silicon device.

What is a Neural Engine?

The Neural Engine is Apple’s brand name for a Neural Processing Unit (NPU) — a dedicated piece of hardware designed to accelerate machine learning models. Apple describes it as “dedicated hardware in its chip architecture for accelerating AI workloads,” per the official Apple Newsroom (company press team). It first appeared in the A11 Bionic chip powering the iPhone 8 and iPhone X, released in 2017.

What is a Neural Engine on an iPhone?

• On iPhones, the Neural Engine handles tasks like Face ID recognition, Animoji tracking, and real-time photo analysis without sending data to the cloud.
• Every iPhone since the A11 chip (iPhone 8, iPhone X, and all later models) includes a Neural Engine.
• Apple’s machine learning research confirms the Neural Engine can run complex models like Transformers efficiently (Apple Machine Learning Research (academic arm)).

The pattern: each generation widens the performance gap, making earlier devices obsolete for on-device AI.

What is the purpose of a Neural Engine?

The Neural Engine exists for one reason: to run AI and machine learning tasks locally, fast, and with low power. According to Apple’s developer documentation, it handles real-time tasks like image recognition, natural language processing, and video analysis (Apple Newsroom (Apple Intelligence launch)). By processing on the chip rather than in the cloud, it keeps user data private and reduces latency.

What is a neural engine used for?

• Image processing: Photos app uses it for face detection, scene recognition, and search.
• Speech and language: Siri dictation, keyboard predictions, and live captions run on the Neural Engine.
• Camera features: Portrait mode, Night mode, and Deep Fusion all lean on Neural Engine computations.

What is Apple Neural Engine used for?

• Apple Intelligence — Apple’s personal intelligence system announced at WWDC on June 10, 2024 — relies heavily on the Neural Engine to run on-device foundation models (Apple Newsroom (WWDC 2024)).
• Apple’s research team published a description of a roughly 3 billion-parameter on-device language model that runs specifically on the Neural Engine (Apple Machine Learning Research (foundation models)).
• For complex requests, Apple Intelligence can fall back to Private Cloud Compute while still protecting privacy (Apple Support (privacy documentation)).

What this means: Apple’s AI strategy hinges on the Neural Engine’s ability to scale, and the gap between on-device and cloud will narrow as chips evolve.

Does the MacBook have the Neural Engine?

All MacBooks and Mac desktops with Apple Silicon — the M1, M1 Pro, M1 Max, M1 Ultra, M2 series, and M3 series — include a 16-core Neural Engine. Intel-based Macs do not have one. That means older Macs bought before late 2020 lack the dedicated ML hardware entirely.

Neural Engine Mac capabilities

• The M1, M2, and M3 families all use a 16-core Neural Engine capable of up to 11 trillion operations per second in the M1 generation and 18 trillion in the M2 generation.
• Apple’s machine learning research shows that applications can use the Neural Engine for energy-efficient, high-throughput inference with Transformers (Apple Machine Learning Research (Transformer deployment)).
• macOS apps like Pixelmator Pro, DaVinci Resolve, and Adobe Lightroom use the Neural Engine for tasks like object masking, upscaling, and subject selection.

The implication: upgrading to Apple Silicon isn’t just about speed — it’s about access to the entire on-device AI ecosystem.

Who makes NPU chips?

Apple designs its own Neural Engine in-house as part of its Apple Silicon strategy. But the broader NPU landscape features several major players. Qualcomm’s Hexagon NPU powers Snapdragon chipsets, Samsung includes its own NPU in Exynos processors, and MediaTek integrates NPUs into its Dimensity line. Google also designs its Tensor Processing Unit (TPU) for Pixel devices.

Apple Neural Engine vs Qualcomm Hexagon

• Apple’s approach: tightly integrated with the rest of the SoC, tuneable through Core ML and Apple’s own neural network framework.
• Qualcomm Hexagon is optimized for Android devices and runs through Qualcomm’s Neural Processing SDK. Benchmarks vary widely, but Apple’s per-core efficiency tends to lead in first-party benchmarks.

MediaTek NPU

• MediaTek’s NPU, found in Dimensity chips, focuses on camera AI tasks, voice recognition, and gaming optimizations. Performance claims from MediaTek exceed 4 trillion ops/sec on their flagship designs.

Samsung NPU

• Samsung integrates NPUs into Exynos chips for Galaxy devices. Their dual-NPU approach handles camera and voice tasks separately for better parallelization.

The catch: Apple’s closed ecosystem may limit third-party NPU adoption, but the hardware trend is clear — every flagship SoC will have an NPU.

What is the strongest Apple CPU?

As of 2024, the strongest Apple CPU in terms of raw compute is the M3 Ultra, which packs up to 32 CPU cores and a 16-core Neural Engine. But the M2 Ultra remains competitive, especially in multi-threaded workloads. The M4 chip, while focused on efficiency, delivers the fastest Neural Engine yet at 38 trillion operations per second — though Apple hasn’t released a direct comparison chart for all generations.

Apple Neural Engine architecture details

• The Neural Engine architecture is a proprietary neural network accelerator consisting of multiple processing cores designed for matrix math and convolution operations.
• Each generation increases the core count and per-core throughput. The A11 had 2 cores, A12 had 8, and the M1 family scaled to 16.
• Apple’s press materials explicitly state the M4 Neural Engine is “60x faster” than the A11’s original iteration (Apple Newsroom (M4 launch)).

The pattern: raw performance jumps are massive, but Apple keeps architectural details proprietary, limiting independent verification.

Is a 16-core Neural Engine good?

A 16-core Neural Engine, standard across M-series chips, handles most current on-device machine learning workloads well. For tasks like photo processing, Siri responses, and Apple Intelligence inference, 16 cores provide good headroom. Benchmarks from Apple indicate that the M1’s 16-core engine can perform up to 11 trillion operations per second, which is enough to run Apple’s 3 billion-parameter language model locally.

Apple Neural Engine vs GPU performance

• The GPU is a general-purpose parallel processor for rendering graphics and running compute shaders. The Neural Engine is specialized for tensor operations — matrix multiply, convolutions, activations. For AI inference, the Neural Engine uses less power and runs faster than the GPU for supported models.
• Apple’s research confirms that Transformers deployed on the Neural Engine achieve “energy-efficient, high-throughput inference” compared to GPU deployment (Apple Machine Learning Research (Neural Engine Transformers)).

Neural Engine programming considerations

• Developers access the Neural Engine indirectly through Apple’s Core ML framework. They cannot program the Neural Engine directly — it’s a closed accelerator.
• Apple announced in 2025 that developers could access the on-device foundation model at the core of Apple Intelligence (Apple Newsroom (2025 developer access)).
• This means third-party apps can now leverage the Neural Engine for custom AI models without writing low-level code.

What this means: the Neural Engine’s efficiency makes it the go-to for on-device AI, but its closed nature limits developer flexibility compared to GPU or CPU compute.

The takeaway: Apple’s claims are well-sourced, but independent validation remains scarce — a common theme in closed hardware ecosystems.

Expert perspectives

“The Neural Engine is dedicated hardware in Apple’s chip architecture for accelerating AI workloads — it’s the engine that makes on-device intelligence possible.”

— Apple Newsroom, M4 chip announcement (Apple Newsroom (official press))

“Apple Intelligence uses on-device processing for the most part — it first tries the device chip and only offloads to Apple’s private cloud when needed.”

— PCMag analysis (PCMag (tech press))

The implication: Apple has bet the entire Apple Intelligence strategy on the Neural Engine’s ability to handle ever-larger models locally. For users considering an upgrade, the choice is clear — if you want on-device AI features, Apple Silicon (M-series or later) is mandatory. For Intel Mac owners, the window is closing fast: either upgrade to Apple Silicon or accept cloud-only AI access.

Frequently asked questions