Researchers reverse-engineer Apple's Neural Engine, exposing an undocumented direct path

A new arXiv paper delivers the most detailed public account yet of the Apple Neural Engine (ANE), the fixed-function matrix accelerator baked into Apple silicon since the A11 and M1 chips. Apple only exposes the hardware to developers through its high-level Core ML framework, but the authors reconstruct what sits beneath it by combining direct measurement on M1 and M5 hardware with static analysis of the private runtime, compiler, kernel driver, and firmware. The result documents the engine’s datapath and performance roofline, the on-disk program format, the weight-compression scheme, and the command protocol the driver uses to talk to the silicon — spanning the A11 through A18 and M1 through M5 families with per-chip target tables.

The most consequential finding is a dispatch route that reaches the ANE directly from ordinary user space, bypassing Core ML entirely. It is callable today but undocumented, unsupported, and fragile across OS versions, so the authors frame it as a tool for measurement and on-device research rather than something to ship in production software, where Core ML remains the sanctioned interface. To keep the work honest, every claim is explicitly tagged as measured, decompile-derived, or predicted, with methodology and open questions recorded.

The significance is twofold: it gives researchers a roofline model and low-level handle for squeezing more out of Apple’s ML accelerator, and it pulls back the curtain on a deliberately opaque piece of hardware. That visibility is useful for performance and security analysis, but the version-fragility of the direct path underscores why Apple keeps it hidden — any reliance on it would break with firmware and OS updates.