Retrospectively Reverse-Engineering Apple's Neural Engine

The author revived a three-year-old effort to reverse-engineer the M1 Apple Neural Engine, detailing its compute cores, datapath, scheduler, memory, and execution model. The ANE contains 16 cores, each with 128 FP16 (or 256 INT8) MAC lanes, totaling 2,048 parallel MAC units, and its surrounding dataflow architecture dictates its effectiveness for CNN and transformer workloads.