Apple rarely saves its biggest silicon news for a Tuesday morning press release. Yet that is exactly how the M6 and the M5 Ultra arrived on August 25: no keynote, no livestream, just two refreshed desktops and a spec sheet that resets the baseline for what a desktop computer can do with AI. The M6, Apple’s first chip built on a 2-nanometre process, now powers a new Mac mini. The M5 Ultra, the company’s first quad-die processor and the most powerful chip it has ever shipped, goes into an updated Mac Studio.
A 2-nanometre first
The M6’s headline feature is the process it is built on. Moving to 2 nanometres packs more transistors into a smaller die, and Apple has spent that budget on a larger 12-core CPU, two more cores than the M5, arranged as two super cores, four performance cores and six efficiency cores. Apple claims the chip delivers the world’s fastest single-threaded performance, up to 1.2 times the multithreaded performance of the M5, and up to 2.4 times the speed of the original M1.
The AI hardware got the bigger overhaul. The M6 introduces a Dual 16-core Neural Engine that Apple says provides up to twice the peak compute of previous generations, and system frameworks can drive both engines simultaneously without developers rewriting their apps. The 12-core GPU carries a Neural Accelerator in every core, lifting peak GPU compute for AI by nearly 30 percent over the M5 and more than eightfold over the M1. Unified memory tops out at 32GB with 170GB/s of bandwidth.
“Built using the cutting-edge 2 nm process, M6 combines a new CPU complex, two additional CPU and GPU cores, a Dual 16-core Neural Engine, and more unified memory bandwidth to power through workloads with amazing energy efficiency,” said Sri Santhanam, Apple’s vice president of Silicon Engineering Group, in the announcement.
Four dies, 512 gigabytes
The M5 Ultra is the stranger achievement. Apple’s UltraFusion interconnect has previously fused two chips into one; this generation joins two dual-die M5 Max chips into a quad-die package, a first for Apple silicon, with more than 4.4TB/s of bandwidth flowing between the dies so the four behave as a single processor.
The resulting numbers read like a workstation vendor’s wish list. Up to 36 CPU cores, split between 12 super cores and 24 performance cores. Up to 80 GPU cores, each with its own Neural Accelerator, which Apple says delivers 4.5 times the peak GPU compute for AI of the M3 Ultra and graphics performance up to 40 percent faster. Most significant of all: up to 512GB of unified memory with 1.2TB/s of bandwidth, 50 percent more than the M3 Ultra offered.
That memory figure is the real story. A model with hundreds of billions of parameters can now sit entirely in local memory on a machine that fits under a monitor. Apple is explicit about the target audience: researchers and professionals who want to run frontier-scale AI models on their own hardware rather than renting time in someone else’s data centre.
The local AI bet
Apple’s frameworks are the other half of the pitch. Core AI, Core ML, Metal and Xcode are wired directly into the new hardware, and developers can fine-tune large models locally, tap Apple’s own Foundation Models, or bring proprietary models of their own. For a company that has watched rivals monetise cloud inference, the strategy is clear enough: make the Mac the machine where AI runs privately, on device, with no per-token bill attached. Apple Intelligence features built for the new chips arrive with macOS 27 this fall.
There is a quieter enterprise angle too. Businesses handling sensitive data have been slow to push it through cloud AI APIs. A desktop that runs very large models offline changes that conversation, particularly in regulated industries where data residency is a hard requirement rather than a preference.
The M6 will make its way into laptops in due course, and the more interesting question is what happens to cloud inference budgets when a machine on a desk can hold a frontier-scale model in memory. The answer will come from developers over the next few months, not from Cupertino. For more coverage of AI hardware and the tools built on it, visit Mylistingo.
