Reverse-engineered account of the Apple Neural Engine

Category: Apple
Status: developing
Sources: arXiv 2606.22283, discussion
Summary: Spencer H. Bryngelson published a guide dated 2026-06-21 documenting the Apple Neural Engine (ANE), the fixed-function matrix accelerator shipped in Apple systems-on-chip since the A11 and M1 generations and normally reachable only through Core ML. Based on direct measurement on M1 and M5 silicon plus static analysis of the private runtime, compiler, kernel driver, and firmware, it documents the datapath and roofline, the dispatch route beneath Core ML, the on-disk program format, the weight-compression scheme, and the command protocol, covering A11 through A18 and M1 through M5 with per-chip target tables. The author labels each claim as measured, decompile-derived, or predicted, and notes the direct user-space route is undocumented, unsupported, and version-fragile.
Why it matters: It is the most detailed public map of an otherwise closed accelerator, useful for on-device ML measurement and research on Apple Silicon, though it is a single-author reverse-engineered account rather than vendor documentation.