Apple unleashes M5 with massive on-device AI gains

For the past few years, the entire tech industry treated the NPU as the holy grail of on-device AI. Apple took a completely different turn with its newly announced M5 processor.

Unveiled on October 15, the M5 chip shifts the AI battleground directly to the graphics processor, overhauling Apple silicon's internal balance of power.

Neural Accelerators in every GPU core

Built on third-generation 3-nanometer technology, the M5 features a redesigned 10-core GPU architecture. The headline feature is simple: every single GPU core now contains its own dedicated Neural Accelerator.

That architectural tweak pays massive dividends on paper. Apple claims the M5 delivers over 4x the peak GPU compute performance for AI compared to M4, and more than 6x compared to M1.

For anyone running heavy on-device models, such as diffusion pipelines in Draw Things or local LLMs through webAI, this architecture bypasses standard compute bottlenecks. Developers can target these units directly via new Tensor APIs in Metal 4.

Traditional graphics performance did not get left behind. With third-generation ray tracing and a rearchitected second-generation dynamic caching system, the GPU delivers up to 30 percent faster general graphics and up to a 45 percent boost in ray-traced workloads over M4.

Wider bandwidth and a Vision Pro refresh

Raw compute means nothing if memory cannot feed it fast enough. M5 upgrades unified memory bandwidth to 153GB/s, a roughly 30 percent bump over M4 and double the bandwidth of M1.

On the CPU side, M5 packs a 10-core setup with four performance cores and six efficiency cores, delivering up to 15 percent faster multithreaded gains over M4. A revamped 16-core Neural Engine handles background system intelligence alongside the CPU and GPU accelerators.

Apple is not waiting around to ship the silicon. The M5 is available immediately for pre-order across the new 14-inch MacBook Pro, iPad Pro, and Apple Vision Pro. In the Vision Pro, the chip drives 10 percent more pixels on the micro-OLED panels at refresh rates up to 120Hz, while accelerating spatial photo conversions and Persona generation.

Why it matters

We think M5 signals a crucial tactical shift for edge AI hardware.

Until now, chipmakers partitioned AI tasks into an isolated NPU while treating GPUs as visual engines. But real-world generative models chew through matrices in ways conventional mobile NPUs struggle to scale. By embedding neural engines straight into GPU cores, Apple is effectively turning the GPU into a full-scale AI matrix engine without sacrificing battery life.

For developers and power users, local diffusion models and on-device code generation will feel substantially faster on day one. Apple Intelligence features like Image Playground will benefit directly, and Metal 4 Tensor APIs will likely accelerate third-party library adoption.

The catch is memory headroom. A 32GB ceiling on the baseline M5 still limits how far power users can push massive 70B parameter models locally. Still, if you are sitting on an M1 or M2 machine, this architectural jump makes local AI execution feel like a native capability rather than an afterthought.