iPhone 17 Pro as a MacBook Pro compute extender—actual benchmarks are in.
Apple embedded a neural accelerator in each A19 Pro GPU core, claiming MacBook Pro-tier compute for GPU-heavy AI tasks. A dev tested this by tethering an iPhone 17 Pro Max to a 24 GB M4 Pro MacBook via 10 Gb/s USB-C and running Qwen3.8-27B (IQ4_XS).
Layer split: Mac handles layers 1–40, iPhone takes 41–64. Results at 256-token prefill batches:
• 16k context: 109 → 157 tokens/sec
• 32k context: 101 → 130 tokens/sec
• 48k context: 87 → 113 tokens/sec
Cold 27k-token session dropped from 245s (Mac alone) to 168s (Mac + iPhone). Greedy token outputs matched perfectly.
Above 64k tokens, the architecture shifts—Mac runs the full model while iPhone stores oldest KV pages and computes attention using GPU matrix units + Neural Engine. Mac alone holds ~64k tokens of 8-bit context; iPhone extends that to ~196k–229k on 17 Pro Max. Tested up to 128k at 8-bit and 140k at 4-bit.
Below 64k, iPhone doesn't accelerate decode—stays at Mac's 25 tokens/sec. Critical constraint: iPhone must stay unlocked and in foreground; iOS suspends GPU work otherwise.
This isn't general-purpose eGPU or arbitrary graphics offload. It's a specific pipeline where iPhone's GPU matrix units process the second half of a large model while the host handles the first half and final logits. The same silicon Apple marketed for on-device AI now delivers measurable prefill throughput and extended context capacity when cabled to a Mac.
One model, one layer split, one cable speed, one thermal envelope—but the numbers are real and reproducible.
Apple embedded a neural accelerator in each A19 Pro GPU core, claiming MacBook Pro-tier compute for GPU-heavy AI tasks. A dev tested this by tethering an iPhone 17 Pro Max to a 24 GB M4 Pro MacBook via 10 Gb/s USB-C and running Qwen3.8-27B (IQ4_XS).
Layer split: Mac handles layers 1–40, iPhone takes 41–64. Results at 256-token prefill batches:
• 16k context: 109 → 157 tokens/sec
• 32k context: 101 → 130 tokens/sec
• 48k context: 87 → 113 tokens/sec
Cold 27k-token session dropped from 245s (Mac alone) to 168s (Mac + iPhone). Greedy token outputs matched perfectly.
Above 64k tokens, the architecture shifts—Mac runs the full model while iPhone stores oldest KV pages and computes attention using GPU matrix units + Neural Engine. Mac alone holds ~64k tokens of 8-bit context; iPhone extends that to ~196k–229k on 17 Pro Max. Tested up to 128k at 8-bit and 140k at 4-bit.
Below 64k, iPhone doesn't accelerate decode—stays at Mac's 25 tokens/sec. Critical constraint: iPhone must stay unlocked and in foreground; iOS suspends GPU work otherwise.
This isn't general-purpose eGPU or arbitrary graphics offload. It's a specific pipeline where iPhone's GPU matrix units process the second half of a large model while the host handles the first half and final logits. The same silicon Apple marketed for on-device AI now delivers measurable prefill throughput and extended context capacity when cabled to a Mac.
One model, one layer split, one cable speed, one thermal envelope—but the numbers are real and reproducible.