User Connects iPhone 17 Pro Max to MacBook Pro to Speed Up AI Model, Boosting Prefill Performance by Up to 44%

Qwen3.8-27B is a capable AI model, but running it on an M4 Pro MacBook Pro with only 24GB of unified memory limits prefill speed. A Reddit user connected an iPhone 17 Pro Max to the Mac via USB-C and used homemade open-source software to split the workload, boosting prefill performance by up to 44%.

The setup splits the model by layers: the Mac handles layers 1 to 40 of each 256-token batch, then streams activation data to the phone, where the iPhone's A19 Pro GPU runs layers 41 to 64. GPU processing on the phone is about 2.4 times faster than without it. The Neural Engine also helps, cutting single-token write time from 279ms to 176ms at 140K context length.

Tests show prefill speed rose from 132 to 177 tokens per second at 8K context (35% gain), from 109 to 157 at 16K (44% gain), and from 101 to 130 at 32K (29% gain). However, the phone does not speed up text generation within 64K context, which still runs entirely on the Mac. The software is open-sourced on GitHub under the name backburner.