Apple Neural Engine: Architecture, Programming, and Performance
25 days ago
- The Apple Neural Engine (ANE) is a fixed-function matrix accelerator used in Apple A11 through A18 and M1 through M5 chips.
- It is accessible through Core ML, but this article reverse-engineers its architecture, programming, and performance via direct measurements and static analysis.
- Documentation includes the datapath, roofline model, dispatch route, compiler, on-disk program format, weight compression, and kernel driver/firmware.
- Direct measurements were performed on M1 and M5 chips, with claims categorized as measured, decompile-derived, or predicted.
- An undocumented user-space callable route is described, intended for research and measurement, not for shipping software, where Core ML remains the supported path.