Retrospectively Reverse-Engineering Apple's Neural Engine

A researcher reverse-engineers Apple's Neural Engine (ANE) from the M1 chip to document its full internal architecture before its apparent demise, as Apple has integrated NPU functionality into GPU cores in the M5. The analysis reveals that the ANE was specifically optimized for dense CNN workloads with predictable data reuse patterns, and its specialized design—rather than just its multiply-accumulate units—determined its effectiveness, assumptions that were broken by the shift to transformer-based models.