Apple's M6 and M5 Ultra Chips Debut a New Core AI Framework for On-Device Model Training
Apple's new M6 and M5 Ultra chips ship alongside Core AI, a new framework for building and deploying AI models on Apple silicon, with M5 Ultra supporting 512GB of unified memory.
Overview
Apple introduced two new Apple silicon chips, the M6 and the M5 Ultra, and debuted a new software framework called Core AI for building, running, and deploying AI models on Apple silicon, according to Apple Newsroom. The M6 ships in a refreshed Mac mini and the M5 Ultra in a new Mac Studio, both available for pre-order the same day the chips were announced, with deliveries beginning September 22, according to Apple Newsroom.
What We Know
M6 is “Apple’s first 2 nm chip,” according to Apple Newsroom, a claim independently reported by 9to5Mac: “The M6 is Apple’s first 2-nanometer chip, which the company says increases transistor density and power efficiency.” The chip carries a 12-core CPU and a 12-core GPU, up from a 10-core CPU and 10-core GPU on M5, MacRumors reported, with a Neural Accelerator built into each GPU core. M6 also introduces a “Dual 16-core Neural Engine” — “the first-ever dual Neural Engine,” according to MacRumors — delivering “up to 2x the peak compute over previous generations,” according to Apple Newsroom. In the Mac mini, M6 supports up to 32GB of unified memory and up to 170GB/s of memory bandwidth.
M5 Ultra is Apple’s “first quad-die architecture,” built by fusing two M5 Max dies with a new generation of Apple’s UltraFusion interconnect, which “delivers more than 4.4TB/s of inter-die bandwidth and over 6x higher connection density,” according to 9to5Mac. The chip scales up to a 36-core CPU with 12 super cores and 24 performance cores, and up to an 80-core GPU that “brings Neural Accelerators to the Ultra chip for the first time,” Apple said in its Mac Studio press release. “With M5 Ultra, Mac Studio achieves up to 4.3x the peak AI compute performance of M3 Ultra and a staggering 9.8x more than M1 Ultra,” Apple said in the same release. The new Mac Studio supports up to 512GB of unified memory and up to 1.2TB/s of memory bandwidth, a spec MacRumors confirmed independently, reporting “the new Mac Studio with the M5 Ultra supports up to 512GB of unified memory and up to 1.2TB/s of memory bandwidth.”
The developer-facing centerpiece of the announcement is Core AI. “Core AI is a brand-new framework for building, running, and deploying AI models on Apple silicon,” Apple said in its Mac Studio press release, adding that it “provides an architecture optimized for Apple silicon, including unified memory, CPU, GPU, and Neural Engine, allowing developers to deploy full-scale LLMs locally and bring their own custom models into their apps.” Alongside Core AI, Apple pointed to its existing open-source MLX framework, saying it “enables developers to run, train, and fine-tune models with exceptional efficiency on Mac,” according to the same release. Apple’s chip-specific press release lists “Core AI, Core ML, Metal, and Xcode” as the developer frameworks and tools that “tap directly into the advanced hardware of both chips,” and says that “with these frameworks and new chips, developers can run and fine-tune large AI models locally on their Mac.”
On pricing, the Mac mini with M6 starts at $899 and the M5 Pro configuration starts at $1,699, according to MacRumors. The Mac Studio with M5 Max starts at $2,499 and the M5 Ultra configuration starts at $5,499, according to Apple Newsroom. Both machines are available for pre-order now, with general availability beginning September 22; the 512GB-memory Mac Studio configuration is “coming in late October,” per the same release.
What We Don’t Know
Apple’s performance comparisons — including the 2x, 4.3x, and other multipliers cited above — come from Apple’s own internal testing using pre-production hardware and “select industry-standard benchmarks,” per the press releases’ footnotes; independent third-party benchmarks of shipping units are not yet available. Apple has not disclosed transistor counts, die size, or which foundry manufactures the 2-nanometer M6, and none of the sourced coverage names a fab partner. It also remains unclear whether Core AI will support model formats or tooling from outside Apple’s own ecosystem, or how it will interoperate with existing third-party inference frameworks on macOS beyond the MLX and Core ML integrations Apple has described.