Home
Technology
Gaming
Entertainment
World & Business
Science
Sports
AI

World & Business

Home
Technology
Gaming
Entertainment
World & Business
Science
Sports
AI

iPhone 17 Pro Max reportedly speeds up AI prompt processing on an M4 Pro MacBook · Digg

Report

iPhone 17 Pro Max reportedly speeds up AI prompt processing on an M4 Pro MacBook

A post describes a developer's USB-C test that raised 16 k-context prefill from 109 to 157 tokens per second.

1 Source, 2h ago, first seen 2h ago

TLDR

A post describes a developer's test of Qwen3.8-27B on a 24 GB M4 Pro MacBook, with an iPhone 17 Pro Max connected by a 10 Gb/s USB-C cable to process some model layers. At 16 k context, the reported prompt-processing rate rose from 109 to 157 tokens per second. The post says longer-context tests reached 128 k at 8-bit and 140 k at 4-bit. Below 64 k, the writing rate remained 25 tokens per second, and the phone had to stay unlocked and in the foreground.

Combined views

—

1 Source, first seen 2h ago

Combined views

—

1 Source, first seen 2h ago

— likes

— likes

— comments

— saves

— reposts

— comments

— saves

— reposts

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Today's Rank

—

Not ranked yet

Today's Rank

—

Not ranked yet

1 Source

@BrianRoemmeleiPhone 17 As A MacBook Pro Accelerator! Apple placed a neural accelerator in each GPU core of the A19 Pro and stated that the result is MacBook Pro levels of compute in an iPhone, suited to GPU-intensive AI workloads. The claim appears in the September 2025 keynote for the iPhone 17 series. Independent measurements now show what that compute does when the phone is attached to a larger machine. A developer running a 24 GB M4 Pro MacBook measured Qwen3.8-27B (IQ4_XS) with and without an iPhone 17 Pro Max connected by a 10 Gb/s USB-C cable. The Mac executes layers 1–40 of each 256-token prefill batch. The phone executes layers 41–64 on its GPU and returns the residual. At an existing 16 k context the combined rate rose from 109 tokens per second to 157; at 32 k from 101 to 130; at 48 k from 87 to 113. Those are the published figures from October 1, 2026. A cold 27 k-token session fell from 245 seconds on stock llama.cpp to 168 seconds with the phone attached. Greedy tokens matched the Mac-only run on the samples tested. Past 64 k tokens the division changes. The Mac runs the full model while the phone stores the oldest key-value pages and computes attention over them, using both the GPU matrix units and the Neural Engine. The Mac alone holds roughly 64 k tokens of 8-bit context beside the model; the phone’s free memory extends that range to an estimated 196 k–229 k on a 17 Pro Max, with end-to-end tests reported to 128 k at 8-bit and 140 k at 4-bit. Below 64 k the phone does not accelerate decode; the measured writing rate stays at the Mac’s 25 tokens per second. The phone must remain in the foreground; iOS suspends GPU work in the background. The numbers are limited to one model, one layer split, one cable speed, and one thermal envelope. They do not describe arbitrary graphics offload or general-purpose external GPU use. They do show a concrete pipeline in which the phone’s GPU matrix units process the second half of a large model while the host supplies the first half and the final logits. The same hardware that Apple described as MacBook-class for on-device AI now contributes measurable prefill throughput and additional context capacity when it is cabled to a Mac. The constraint remains that the phone stays unlocked and visible.2h