Chips Powering Apple Intelligence in iOS 18

Apple Intelligence in iOS 18 relies on specialized on-device silicon to run generative AI models locally, safeguarding user privacy while minimizing latency. The system depends primarily on Apple's dedicated Neural Engine (NPU), backed by a unified memory architecture and supplemental acceleration from custom GPU and CPU cores. This article details the specific hardware components and silicon requirements that make local Apple Intelligence execution possible.

The Apple Neural Engine (NPU)

The primary component handling on-device AI tasks is the Apple Neural Engine (ANE). Built explicitly for high-throughput, energy-efficient matrix multiplication and tensor operations, the ANE accelerates machine learning algorithms, including large language models (LLMs) and diffusion models.

In iOS 18, Apple Intelligence requires a 16-core Neural Engine capable of processing high operations-per-second thresholds:

  • A17 Pro: Features a 16-core Neural Engine capable of up to 35 trillion operations per second (TOPS), nearly twice as fast as the A16 Bionic.
  • A18 and A18 Pro: Feature updated 16-core Neural Engines with increased memory bandwidth, delivering up to 35 TOPS with optimized efficiency for generative foundation models.
  • M-Series Silicon (M1 through M4): Feature 16-core (or up to 38-core in select variants) Neural Engines delivering between 11 TOPS (M1) and 38 TOPS (M4).

Supported System-on-Chips (SoCs)

Because of the heavy compute demands of modern generative AI, only specific chips satisfy the baseline hardware profile for iOS 18 on-device intelligence:

  • iPhone: Exclusively restricted to the A17 Pro (found in the iPhone 15 Pro and iPhone 15 Pro Max) and the A18/A18 Pro chips (powering the iPhone 16 family). Older A-series chips lack the required throughput and memory bandwidth.
  • iPad and Mac: Any device powered by an M1, M2, M3, or M4 chip.

Unified Memory Architecture (UMA)

Neural processing hardware cannot function without sufficient memory throughput and capacity. Apple Intelligence establishes a strict minimum requirement of 8GB of unified memory.

Apple's Unified Memory Architecture allows the CPU, GPU, and Neural Engine to access a single, shared pool of high-bandwidth memory without duplicating data across separate memory buses. This zero-copy architecture allows on-device models—typically quantized parameters compressed to fit within small footprints—to reside directly in memory where the Neural Engine can immediately evaluate tokens.

Assisting Roles: GPU and CPU

While the Neural Engine handles the bulk of inference, other on-die processors assist depending on the workload:

  • Apple GPU: Works in tandem with the ANE using Metal Performance Shaders (MPS) Graph to accelerate visual workloads, image generation, and diffusion models that benefit from high parallel floating-point processing.
  • Apple CPU: Manages token orchestration, task scheduling, input preprocessing, and speculative decoding to keep the Neural Engine saturated with data.