# Apple Neural Engine Apple's dedicated [[Neural Processing Unit (NPU)]] integrated into A-series and M-series chips. First shipped in the A11 Bionic (iPhone 8/X, 2017) and now present in every iPhone, iPad, and Mac. The most widely deployed NPU in the world. Also called: Neural Engine, ANE. ## Generations | Year | Chip | TOPS | |---|---|---| | 2017 | A11 | 0.6 | | 2020 | A14 / M1 | 11 | | 2023 | A17 Pro / M3 | 35 | | 2024 | A18 Pro / M4 | 38 | Performance has grown ~60× across the lineage; capability has grown faster than raw TOPS thanks to architectural improvements (sparsity, mixed precision). ## What It Powers - Photos: object recognition, scene understanding, face detection - Live Text and Visual Look Up - Siri on-device speech recognition - FaceID and TouchID - Camera computational photography (Deep Fusion, Smart HDR, Cinematic mode) - [[Apple Intelligence]] (since 2024) ## Programming Model Developers don't program the Neural Engine directly. Access is via [[Apple Core ML]]: 1. Convert model to Core ML format (`.mlmodel` / `.mlpackage`) 2. Core ML decides at runtime where to execute: ANE, GPU, or CPU 3. Optimal models target the ANE-friendly subset of operations Some models can't run on the ANE (unsupported ops, dynamic shapes) and fall back to GPU/CPU. ## Trade-offs **Strengths:** extreme efficiency (perf/watt), low-latency, ubiquitous on Apple devices, free for all developers **Limitations:** Apple-only, opaque (no public spec), supports a curated op subset, optimization requires model surgery ## Browser Implications When Safari implements [[WebNN API]], expect it to expose Apple Neural Engine via Core ML as the underlying execution provider — making on-device browser ML competitive with native apps on Apple devices. ## References - https://en.wikipedia.org/wiki/Apple_silicon#Neural_Engine ## Related - [[Neural Processing Unit (NPU)]] - [[Apple Core ML]] - [[Apple Intelligence]] - [[On-Device Machine Learning]] - [[Edge AI]] - [[WebNN API]] - [[AI Inference]]