@Underfox3: In this paper is presented a comprehensive guide to the Apple Neural Engine, based on direct measurement on Apple silic…
Summary
A comprehensive reverse-engineering guide to the Apple Neural Engine, detailing its architecture, programming interface, and performance across A11–A18 and M1–M5 chips, based on direct measurements and static analysis of the private runtime, compiler, and firmware.
View Cached Full Text
Cached at: 06/29/26, 06:24 AM
In this paper is presented a comprehensive guide to the Apple Neural Engine, based on direct measurement on Apple silicon and static analysis of the private runtime, compiler, kernel driver, and firmware.
https://t.co/eDdqDtT2yq https://t.co/pXMIxaWPVk
Apple Neural Engine: Architecture, Programming, and Performance
Source: https://arxiv.org/abs/2606.22283 View PDF
Abstract:The Apple Neural Engine (ANE) is the fixed-function matrix accelerator that has shipped in Apple systems-on-chip since the A11-class iPhone and iPad chips and the M1-class Mac chips, exposed to applications only through the Core ML model framework. This guide reports a reverse-engineered account of the engine, based on direct measurement on Apple silicon and static analysis of the private runtime, compiler, kernel driver, and firmware. It documents the datapath and the roofline that bound the engine’s throughput and energy, the dispatch route that reaches it below Core ML, the compiler and on-disk program format, the weight-compression scheme, and the kernel driver, firmware, and command protocol beneath them. The account covers the A11 through A18 and M1 through M5 families, with per-chip target tables and an operation-by-device matrix; the direct measurements are on the M1 and M5. Claims are labeled as measured, decompile-derived, or predicted, and the methodology and open questions are recorded. The direct route is callable from ordinary user space but remains undocumented, unsupported, and version-fragile; it is intended for measurement, research, and on-device work, not for shipping software, where Core ML remains the supported path.
Submission history
From: Spencer Bryngelson [view email] **[v1]**Sun, 21 Jun 2026 00:17:34 UTC (407 KB)
Similar Articles
Apple Neural Engine: Architecture, Programming, and Performance
This paper presents a reverse-engineered account of Apple Neural Engine (ANE) architecture, programming, and performance across A11-A18 and M1-M5 chips, based on direct measurements and static analysis.
@0x0SojalSec: Apple hid 15.8 TFLOPS of raw AI power in every M4 Mac & iPhone. They only let you use the Neural Engine for inference. …
A developer reverse-engineered Apple's private APIs to enable training neural networks directly on the Apple Neural Engine (ANE) in M4 Macs and iPhones, bypassing CoreML and GPU. The project demonstrates that ANE hardware is capable of training, though with limitations like low utilization and CPU fallbacks for some operations.
Apple announced new on device inference engine for Apple Silicon
Apple announced CoreAI, a new on-device inference engine for Apple Silicon at WWDC, replacing CoreML and supporting larger models up to 20B parameters via optimized inference, with a focus on phones and tablets.
Apple M1 Chip Deep-Dive
This article uses high-resolution micrographs to deeply analyze the internal structure, transistor layout, memory interface, and major functional blocks of the Apple M1 chip, showcasing its revolutionary breakthrough in performance and efficiency.
@akshay_pachaar: Apple finally did it. Its new framework, Core AI, runs models entirely on Apple silicon, so inference happens on the us…
Apple released Core AI, a new framework that runs AI models entirely on Apple silicon devices (iPhone, iPad, Mac, Vision Pro) with zero server calls. It includes a memory-safe Swift API, model export recipes for PyTorch, an optimizer, and debugging tools, supporting models like Qwen, Mistral, and SAM3.