Apple puts Neural Accelerators into every GPU core

Running large models locally on a laptop used to mean choosing between molten aluminum, terrible battery life, and sluggish prompt processing. Apple’s latest answer is basically to weld two dies together and sprinkle neural hardware across the entire GPU.

Apple introduced the 14-inch and 16-inch MacBook Pro powered by the all-new M5 Pro and M5 Max chips. The new machines lean aggressively into on-device artificial intelligence, packing redesigned silicon, doubled starting storage, and a new custom wireless chip.

Neural engines in every GPU corner

The headline shift here is Apple's new Fusion Architecture, an approach that fuses two distinct dies into a single high-performance system-on-chip.

Under the hood, both chips boast an up-to-18-core CPU layout featuring 6 super cores and 12 performance cores, offering a 30 percent boost over the previous generation. But the real star is the graphics architecture. Apple embedded a dedicated Neural Accelerator inside every single GPU core.

That hardware tweak delivers up to 4x faster prompt processing for large language models compared to the M4 Pro and Max, and up to 8x the image generation speed of original M1 chips. Raw graphics performance is also up to 50 percent faster, backed by memory bandwidth reaching 307GB/s on M5 Pro (up to 64GB unified memory) and 614GB/s on M5 Max (up to 128GB unified memory).

Apple also slipped in its own custom N1 chip, finally delivering Wi-Fi 7 and Bluetooth 6 support alongside up to 24 hours of battery life.

Double the base storage and 14.5GB/s speeds

Apple has finally addressed one of the longest-running complaints about its pro laptops: stingy entry-level storage.

The M5 Pro MacBook Pro now starts with 1TB of onboard storage, while the M5 Max jumps to a 2TB baseline. Even the entry 14-inch M5 model now begins at 1TB.

Speed got a massive overhaul too. The new internal SSDs deliver up to 14.5GB/s in read and write speeds, doubling the throughput of the previous lineup. Pre-orders open March 4, with shipments starting March 11.

Why it matters

Apple’s decision to integrate Neural Accelerators directly into individual GPU cores looks much smarter than merely cranking up a standalone NPU.

Standalone NPUs often hit walls when juggling massive matrix operations and prompt digestion for large local weights. By baking neural math directly into GPU units and feeding them with up to 614GB/s of unified bandwidth, Apple has essentially turned a lap-friendly laptop into a self-contained AI workstation.

For engineers, researchers, and creators tinkering with local models, upgrading from M1 or M2 makes absolute sense now. Cutting prompt processing times by up to 8x transforms local inference from a frustrating experiment into a genuine workflow.

That said, if your day mostly consists of basic compiling, office productivity, or web work, this machine is staggering overkill. You are paying heavily for extreme silicon real estate and high-bandwidth unified RAM you will likely never saturate. But in the broader race for the AI PC era, Apple just forced everyone else back to the drawing board on silicon packaging and memory architecture.