Applications

First Snapdragon with INT4 support. The precision floor dropped, which is what later enabled on-device LLM work.

Qualcomm’s peak disclosure: an explicit parameter ceiling and token rate against a named model. Neither successor repeated it.

A wider tensor accelerator for camera and voice AI, with no transformer or generative capability in view.

The first A-series where Apple’s headline AI number describes the GPU rather than the Neural Engine.

Apple’s first 2nm phone chip and its first dual-NPU design, with 50 per cent more memory bandwidth, which matters more for model…

The genuine change was memory bandwidth, which Apple tied directly to Apple Intelligence latency rather than to Neural Engine throughput.

The architectural shift: AI matrix acceleration moves into every GPU core rather than the NPU carrying generative work alone.

The Neural Engine was explicitly re-tuned for large generative models, but Apple benchmarked it against the A16, not the A17 Pro.

A marginal Neural Engine step. The real A16 story was imaging and display, not AI.

The industry’s first 3nm phone chip, and the first A-series for which Apple published no operations-per-second figure at all.