Eileen Yoon retrospectively reverse-engineers Apple's Neural Engine (ANE) architecture after the M5 chip folded ANE cores into GPU, signaling the end of standalone NPUs. She maps the full internal architecture of the M1 ANE — compute cores, datapath, scheduler, memory, and execution model — to understand Apple's design assumptions about ML workloads from the A11 Bionic onward.
Background
Apple first introduced its Neural Engine in the A11 Bionic (2017) as a dedicated NPU for CNN workloads. By 2025, the M5 chip consolidated ANE functionality into GPU cores, reflecting the industry shift from CNN-centric NPUs to GPU-based transformer inference.
- Source
- Lobsters
- Published
- Sep 12, 2026 at 01:23 AM
- Score
- 6.0 / 10