Apple Neural Engine is a series of AI accelerators (NPU) designed by Apple, integrated directly as a block inside A- and M-series SoCs. The first version appeared in September 2017 in the A11 Bionic chip (iPhone 8, iPhone 8 Plus, iPhone X). In 2020 the Neural Engine also arrived in Mac M-series chips, starting with the M1. Every A-series and M-series chip released after 2017 includes a Neural Engine.
The Neural Engine's architecture descends directly from work on the discontinued Project Titan โ Apple's self-driving car program. Ruslan Salakhutdinov (hired in 2017 from Carnegie Mellon as director of AI) designed a chip capable of processing car sensor data in real time, on-device, without sending sensitive data to the cloud. The car never made it to production, but the compute architecture built at that time became the foundation of Apple's on-device AI strategy.
Performance grows with each generation. In the M3 chip the Neural Engine reaches 18 trillion operations per second (TOPS), and in M4 โ 38 TOPS (a more than twofold increase). Another major leap is announced for the M7 Ultra server chip (planned for H1 2027), which is set to build on a substantially expanded Neural Engine and support up to 1.5 TB of RAM.
The Neural Engine powers key AI features across Apple's ecosystem: Face ID facial recognition, the Siri voice assistant, computational photography (Smart HDR, Night Mode), AR effects, and โ from 2024 โ the full Apple Intelligence suite (Image Playground, Writing Tools, upgraded Siri). Developers access it through the Core ML framework, which automatically routes ML models to the most efficient accelerator (Neural Engine, GPU or CPU).
Key architectural advantages: high energy efficiency (critical on mobile devices), on-device processing (biometric and voice data does not leave the device, strengthening privacy) and tight integration with the SoC's unified memory โ CPU, GPU and Neural Engine share the same RAM pool, eliminating the data-transfer bottleneck present in traditional x86 architectures with separate VRAM.
AI SoC / Edge AI SoC ยท serves as: AI acceleration, AI Inference, Compute, Computer vision processing.
Which group Apple Neural Engine belongs to and how it is built
Covers integrated System-on-Chip circuits used as central processing units in embedded systems, robotics, AIoT, and edge AI devices. In this schema, parentCategory is set to 'other' because the parentCategory enum does not include a compute category.
Construction class for NPU (Neural Processing Unit) components implemented as a block inside a larger SoC. Unlike standalone AI accelerator cards (data-center class) or dedicated NPU chips, an on-die NPU block shares silicon with the CPU and GPU, uses the SoC package's unified memory, and is only sold as part of a parent product rather than as a separate component. Typical examples: Apple Neural Engine in A/M-series, Qualcomm Hexagon NPU, Google Tensor TPU (mobile).
Other hardware parts related to Apple Neural Engine
