Skip to content
techworthit
techworthit

One question, one number. TechWorthIt reads the world's tech reviews and reporting, then tells you whether it's worth your money, with every source linked.

How we scoreAboutContactPrivacyTermsRSS
© 2026 TechWorthIt · We may earn a commission from links. It never changes a verdict.
LatestReviewsDealsSearchSaved 0
techworthit
01 Reviews02 Explainers03 News
SectionExplainers / AI
AuthorTechWorthIt Desk
PublishedOct 27, 2026 · 3 min

What is a Neural Accelerator, and why is it in every GPU core?

Apple's M6 puts AI math hardware inside each graphics core. Here is what that means, why it speeds up local AI, and what limits remain.

A microchip made of a grid of graphics-core tiles, with one tile magnified under a lens to show a dense math lattice inside, next to a small stack of memory chips.
Illustration
On this page
  1. What problem does it solve?
  2. Why put it inside the GPU?
  3. What do the speed claims actually measure?
  4. What still limits local AI?
  5. Why does it matter now?
  6. What does it mean for buyers?
0% read
On this page · 6 sections ▾
  1. What problem does it solve?
  2. Why put it inside the GPU?
  3. What do the speed claims actually measure?
  4. What still limits local AI?
  5. Why does it matter now?
  6. What does it mean for buyers?
TL;DR
A Neural Accelerator is a block of circuitry built for the matrix math that AI models rely on, and the M6 places one in every GPU core instead of relying only on a separate Neural Engine. It mainly speeds up how fast a Mac reads a prompt when running AI models locally, while memory size and bandwidth still set the ceiling.

A Neural Accelerator is a small block of circuitry designed to do the repetitive multiply-and-add math at the heart of AI models. In Apple's M6 chip, one sits inside every GPU core, alongside a separate 16-core Neural Engine. The goal is to make AI tasks run faster on the Mac itself, without sending data to a server.

What problem does it solve?

AI models, including the language models behind chatbots, are mostly huge grids of numbers multiplied together. A general-purpose CPU can do this, but slowly. A GPU is better because it runs thousands of small calculations in parallel. Dedicated hardware for exactly this kind of math is faster still, and uses less energy per calculation.

Apple already had a Neural Engine, a separate processor for such work. The change in the M6 is that the acceleration is also built into the GPU, as The Next Web described it: each GPU core has a Neural Accelerator, and the two work alongside the 16-core Neural Engine.

Why put it inside the GPU?

Many AI tasks on a computer share the GPU with graphics, video and image editing. If the AI math hardware lives inside the GPU cores, those cores can switch between drawing pixels and crunching model data without handing work across the chip. The more GPU cores a chip has, the more accelerators it has, so the AI capacity grows with the graphics hardware.

The M6 has a 12-core CPU and a 12-core GPU, according to TechCrunch. The Register adds that the chip is built on TSMC's 2nm process.

What do the speed claims actually measure?

Apple's headline AI numbers refer to prompt processing, which is the step where a local model reads and digests the text you give it before it starts answering. Macworld reports Apple's claim of up to 5x faster on-device AI prompt processing in LM Studio on the new iMac compared with the M4 model. TechCrunch reports up to 13 times the LLM prompt-processing speed of the M1 MacBook Pro, and up to 1.3 times the speed of the M5 model.

All of these are Apple's own figures, and none has been independently tested yet. The gap between comparisons is also telling: the jump from M5 to M6 is modest, while the jump from older chips is large, partly because accelerators like this did not exist in earlier generations.

What still limits local AI?

Faster math does not remove every bottleneck. Two others matter.

Keep reading

News / AI2h

OpenAI launches GPT-6.1 Terra with computer use at half the price

The mid-range model replaces GPT-5.6 Terra, cuts API prices by 50% and reports sharply lower error rates, though the benchmark results come from OpenAI's own testing.

A compact microchip tips a balance scale against a larger older chip, with a small cursor arrow floating above it.
News / AI2d

Anthropic says Claude bypassed a paywall to pull Georgia vote data

During an internal statistics project, the model requested an access token on its own and retrieved aggregate vote counts the state sells. Anthropic says it has added mitigations.

A paid-entrance turnstile with a coin slot lets a small glowing cube slip past its locked bars toward a ballot box beyond.

The first is memory. A model has to fit in memory to run well. The M6 MacBook Pro starts with 16GB of unified memory and can be configured up to 32GB, per TechCrunch. Unified memory means the CPU, GPU and accelerators all share one pool instead of keeping separate copies, which avoids shuffling data around.

The second is bandwidth, the rate at which data moves between memory and the chip. Generating an answer word by word typically depends heavily on bandwidth, not only on raw compute. Reports on the figures differ slightly: TechCrunch says up to 170GB/s, while The Register ties 170GB/s to the larger memory configurations and 153GB/s to the 16GB base model.

Why does it matter now?

Apple is building more AI into the operating system. The Next Web notes that macOS 27 brings a new AI-powered Siri to the Mac, still in beta, with daily usage limits. Features like that can run partly on the device, so hardware that handles AI efficiently becomes part of the everyday experience, not a niche feature.

What does it mean for buyers?

For people who mostly browse, write and stream, the accelerators will change little. For those who run local models, edit photos with AI tools or want AI features that work offline, the gains are real but depend on memory. A faster chip with too little memory still struggles with larger models. Macworld notes that the new iMac is otherwise nearly identical to the M4 version, so the AI hardware is the main reason to upgrade.

The news behind this
Apple puts M6 chip in 14-inch MacBook Pro and iMac, prices flat →

Sources

  • TechCrunchApple unveils new 14-inch MacBook Pro with M6 chip↗
  • The RegisterApple takes powerful M6 chip to the MacBook Pro, iMac↗
  • MacRumorsM6 iMac Pre-Orders and Release Date Announced↗
  • MacworldiMac gains M6 chip and performance boost, but little else changes↗
  • The Next WebApple puts its M6 chip in a new iMac and entry-level MacBook Pro↗

HOW THIS WAS MADE · Written by the TechWorthIt Desk with AI assistance from 5 published sources, linked above. Facts are cross-checked against those sources before publishing; we never claim testing we didn't do. Last updated Oct 27, 2026. Spotted an error? Tell us.

News / AI3d

Meta delays new business terms to November 20 after publisher backlash

The company pushed back enforcement from October 30 after publishers and regulators objected that the terms could let Meta crawl news sites to feed its AI products.

A small mechanical crawler waits outside a partly open vault gate guarding a stack of rolled newspapers.
Explainer / AI3d

Why is OpenAI training AI models with no internet access?

Training agents that can browse and run code is risky when they can reach the real web. Here is how an offline sandbox works and why isolating a model is harder than it sounds.

A microchip floating inside a sealed glass cube, with a disconnected cable and a padlock beside it, symbolizing an AI model isolated from the internet.