A Reddit user going by StayLameBro has hacked together a surprisingly practical solution: an iPhone can act as an external graphics card for a MacBook Pro equipped with an M4 chip. By tethering the phone via a standard 10 Gbps USB‑C cable, the handset’s GPU picks up part of the heavy lifting required for running large language models, effectively extending the laptop’s graphical horsepower without buying a dedicated enclosure.
The setup splits a model—specifically the 27‑billion‑parameter Qwen 3.8—in half. The MacBook handles the first 40 layers of each token batch, while the iPhone’s GPU takes over the remaining layers. When a prompt exceeds 64,000 tokens, the laptop runs every layer itself, and the iPhone merely manages the key‑value caches that store label‑data pairings. The Mac also pulls an 8‑bit context from the phone’s available RAM, keeping its own memory footprint low, and the device helps prefetch inputs beyond the 512‑token mark.
Real‑world testing shows clear benefits. For a 2,000‑token prefill into an existing agent session, the combined Mac‑iPhone rig runs 29 % to 44 % faster than the Mac alone. In a fresh 27,000‑token agent run, speed gains climb to 36 % with the forked llama.cpp fork and reach up to 46 % with the stock implementation.
All the code, configuration steps, and benchmark scripts are openly hosted on GitHub, making it easy for anyone with an iPhone 15 Pro (or newer) to replicate the experiment. The creator notes that an iPhone 18 Pro would provide “a lot more” headroom, implying that newer silicon could push performance even higher.
**What are the downsides?**
The iPhone isn’t a universal fix. Write speeds stay unchanged for prompts under 64,000 tokens, as those tasks remain entirely on the computer. Additionally, the solution hinges on a high‑quality USB‑C connection and a phone with sufficient GPU horsepower; older or non‑Pro models simply won’t cut it.
In short, repurposing a recent iPhone as an external GPU offers a cheap, flexible way to accelerate AI workloads on a Mac, though users should stay aware of its limits and ensure they have the right hardware to make it work reliably.
