NPU Stacks

NPUs (Neural Processing Units) live alongside the CPU and GPU on every modern phone SoC. Each vendor ships its own programming model — Qualcomm Hexagon via QNN (Snapdragon phones, robots, drones), Apple Neural Engine via Core ML (every iPhone since 2017), MediaTek APU, AMD XDNA, ARM Ethos. This module covers the two stacks that ship the most production inference today: Hexagon and ANE.

0 / 2 lessons~30 min total

Module capstone — build it

Profile a model on Hexagon AND ANE — and read the differences

Same model, two NPUs, two profiler tools. The artifact is the side-by-side trace that explains why an op runs at 100 GOPS on one chip and 5 GOPS on the other.

AdvancedOne focused weekend (~12 h)Runs on your phone

Take a small CNN (MobileNetV3) or a tiny transformer (Phi-3.5-mini-128k). Convert to QNN (Hexagon) and Core ML (ANE). Profile each on real hardware: QNN's `qnn-profile-viewer` for Hexagon, Xcode Instruments → Core ML for ANE. Identify three ops that run faster on one NPU than the other; explain why.

You walk away with

A side-by-side NPU trace report — the kind of artifact mobile-ML interviewers ask for by name
Felt knowledge of how Hexagon HVX vectors and ANE matrix tiles differ in practice
A reusable conversion + profiling pipeline you can re-run on any new mobile model

Tools you'll use

QNN SDK 2.20+
Core ML 7+ / coremltools 7.2+
Snapdragon 8 Gen 3 device + Xcode 15+ for ANE
MobileNetV3 or Phi-3.5-mini reference