Apple's new M6 and M5 Ultra bring frontier AI to the desk

Apple just unleashed the next era of its desktop silicon, dropping the 2nm M6 chip and a monstrous M5 Ultra simultaneously.

The M6 debuts inside the refreshed Mac mini, while the M5 Ultra powers the new Mac Studio, targeting creators and AI developers who need serious horsepower without renting a cloud cluster.

M6 brings 2nm silicon and dual neural engines

The M6 is Apple’s very first chip built on a cutting-edge 2nm process. It packs a 12-core CPU layout featuring 2 super cores, 4 performance cores, and 6 efficiency cores, delivering the fastest single-threaded performance in the world alongside up to 2.4x faster multithreaded performance than M1.

Graphics and AI get a major upgrade too. The 12-core GPU features dedicated Neural Accelerators inside each core, pushing peak GPU compute for AI up by 30 percent over M5 and over 8x compared to M1.

Apple also introduced a Dual 16-core Neural Engine that macOS can tap simultaneously, doubling peak compute for on-device AI tasks, backed by up to 32GB of unified memory running at 170GB/s.

M5 Ultra: A quad-die powerhouse with 512GB memory

If the M6 is built for the masses, the M5 Ultra is designed for AI researchers running frontier models locally. For the first time, Apple used UltraFusion to link two dual-die M5 Max chips into a unified quad-die architecture, boasting over 4.4TB/s of inter-die bandwidth.

The specs on this desktop beast are wild: up to a 36-core CPU, up to an 80-core GPU, and a 32-core Neural Engine. GPU peak compute for AI jumps up to 4.5x over M3 Ultra.

The killer feature is unified memory: up to 512GB with a massive 1.2TB/s of bandwidth, up to 50 percent higher than M3 Ultra. That bandwidth and capacity allow developers to load huge frontier models with hundreds of billions of parameters entirely into local VRAM.

Why it matters

Apple is turning the desktop Mac into the ultimate personal AI workstation. For developers and AI researchers, the ability to run a 500B+ parameter model entirely offline on a Mac Studio with 512GB of unified memory is a massive workflow unlock that avoids expensive cloud API fees and data privacy headaches.

For everyday consumers, the M6 Mac mini looks like the clear sweet spot. Anyone still holding onto an M1 Mac gets over double the multithreaded performance, hardware ray tracing, and a machine ready for agentic AI tasks right out of the box.

Strategically, Apple is leaning hard into its unified memory moat. While PC workstations scramble to stitch together multi-GPU setups with tricky power and PCIe lane constraints, Apple's quad-die UltraFusion treats massive memory pools as a single coherent block. The catch remains price and raw ecosystem lock-in: top-tier Mac Studios will cost a fortune, and despite huge inference gains, the heavy lifting of global foundation model training still belongs to CUDA.