Apple debuts M5 Ultra with an 80-core GPU and 512GB of memory

Apple debuted two processors on August 25, 2026: the M5 Ultra in a refreshed Mac Studio, and the M6 in a new Mac mini. The M5 Ultra is the company’s first quad-die M-series design, and Apple calls it its most powerful chip ever. Both machines are aimed squarely at people running AI models on their own hardware, which is a shift in who these desktops are for.

The mechanism is UltraFusion, the interconnect Apple uses to bond dies into one system on a chip. Here it connects two dual-die M5 Max chips, which is a first for Apple silicon. Apple’s release puts inter-die bandwidth at over 4.4TB/s and connection density at over 6x, and that is what lets the four dies behave as a single processor.

What that buys is an up-to-36-core CPU, split into 12 super cores and 24 performance cores, plus a GPU with up to 80 cores. Every GPU core carries a Neural Accelerator. The memory side matters more for anyone running an LLM on their own machine, because model size is usually the binding constraint.

Spec	M6 (Mac mini)	M5 Ultra (Mac Studio)
CPU	12 cores: 2 super, 4 performance, 6 efficiency	Up to 36 cores: 12 super, 24 performance
GPU	12 cores, Neural Accelerator in each	Up to 80 cores, Neural Accelerator in each
Neural Engine	Dual 16-core	32-core
Max unified memory	32GB	512GB
Memory bandwidth	Up to 170GB/s	1.2TB/s
Apple’s AI compute claim	Nearly 30% more peak GPU compute for AI vs M5	Up to 4.5x peak GPU compute for AI vs M3 Ultra

That 1.2TB/s figure is 50 percent higher than the M3 Ultra, and the 512GB pool is the number Apple keeps returning to. It lets users “store huge datasets entirely in local memory, increase the tokens-per-second speed, and run huge LLMs with hundreds of billions of parameters entirely on device,” the company says. The CPU gains are more modest, though, at up to 1.25x single-threaded and 1.3x multithreaded against the M3 Ultra.

And for the ultimate desktop performance and the ability to run massive AI models, M5 Ultra features a massive GPU, now with Neural Accelerators, and more unified memory bandwidth, pushing the boundaries of what a desktop can do.

Sri Santhanam, Apple vice president of Silicon Engineering Group, via Apple Newsroom

Pricing is where the ambition meets a wall. Ars Technica reports that the Mac mini with M6 starts at $899 with 16GB of memory, while Mac Studio with M5 Max starts at $2,499 and M5 Ultra configurations start at $5,499. Preorders opened on announcement day, both ship on September 22, and the 512GB Studio configuration slips to late October. Engadget notes that with memory priced the way it is now, a loaded M5 Ultra Studio could reach five-figure territory.

The refresh follows behaviour Apple did not design for. Ars Technica’s Samuel Axon points out that macOS 26.2, which shipped last December, enabled low-latency communication between Thunderbolt 5 hosts for distributed AI inference using MLX. Since then, hobbyists and researchers have been daisy-chaining Mac minis and Mac Studios, so that they can run models far larger than any single mass-market machine holds, whichever local runtime they picked. That is the same calculation behind choosing a hosted API or your own hardware, and Ars frames token costs from frontier models as the thing pushing developers toward local inference.

There aren’t any verifiable benchmarks to work from yet.

Samuel Axon, Ars Technica

That caveat is the honest reading of every number above. Apple’s own footnotes say testing was conducted by Apple in August 2026, and the M5 Ultra comparisons ran on a preproduction Mac Studio with 256GB of memory, not the 512GB configuration the marketing leans on. So nobody outside Apple has measured tokens per second on this silicon yet, which is the one number that would settle the pitch.

The M6, Apple’s first 2 nm chip, has its own asterisk. Engadget, citing Bloomberg’s Mark Gurman, reported that Apple plans no M6 Pro or M6 Max at all, and intends to skip to the M7 by mid-2027 because of AI processing gains in that chip. If that holds, then the 32GB memory ceiling on the M6 is the ceiling for a while, because no bigger sibling is coming.

TechCrunch reads the launch as Apple playing to its remaining strength, given that the long-awaited Siri upgrade runs on Google’s Gemini. On-device processing is territory Apple already owns, so it’s a plausible place to compete, but the case still rests on unaudited figures. The number worth waiting for is an independent tokens-per-second measurement on a 512GB Studio in late October.