Apple's M6 arrives on 2nm silicon, bringing a major AI compute leap to Mac mini and Mac Studio

Apple on August 25 unveiled two new chips — the M6 and the M5 Ultra — alongside redesigned Mac mini and Mac Studio systems, marking its first step onto 2-nanometer silicon and its boldest push yet into on-device AI compute.

M6: Apple's first 2nm chip

The M6 debuts in the new Mac mini and is the first Apple chip manufactured on a 2nm process. Its 12-core CPU (two super cores, four performance cores, six efficiency cores) delivers up to 1.2x faster multithreaded performance than M5 and is 2.4x faster than M1. Single-threaded performance is once again claimed to be the fastest available on any desktop chip.

For AI workloads, the M6 carries a dual 16-core Neural Engine — double the peak compute of prior generations — and a 12-core GPU with Neural Accelerators embedded in each core, yielding a 30% increase in peak GPU AI compute over M5. Unified memory tops out at 32GB with 170GB/s bandwidth, a 10% bump over M5 and 2.5x over M1.

M5 Ultra: the quad-die Mac Studio

The M5 Ultra, powering the new Mac Studio, takes a different architectural path. Apple linked two dual-die M5 Max chips using its UltraFusion interconnect — the first time a quad-die design has appeared in the M-series lineup. The result is a chip with up to 36 CPU cores, an 80-core GPU, and a 32-core Neural Engine that delivers 4.5x the peak GPU AI compute of the M3 Ultra and 6x that of M1 Ultra.

The headlining spec is memory: the M5 Ultra supports up to 512GB of unified memory with 1.2TB/s of bandwidth — 50% more than the M3 Ultra — and inter-die bandwidth exceeding 4.4TB/s. For machine learning researchers and video production studios running large local models, the 512GB ceiling lifts a real constraint that has pushed some workloads off the desktop and onto cloud infrastructure.

What this means for AI on the desktop

The dual Neural Engine in M6 and the memory headroom in M5 Ultra together signal Apple's intent to make the Mac a serious platform for on-device inference, not just a companion to cloud APIs. At 170GB/s, the M6 Mac mini can comfortably run 7B-to-13B parameter models at full precision. The M5 Ultra's 512GB at 1.2TB/s opens the door to 70B-class models — previously the exclusive domain of multi-GPU server rigs — on a machine that sits quietly on a desk.

Both systems are available now from Apple's online store, though pricing was not disclosed in the announcement. As reported by Apple Newsroom, both chips debuted on August 25, 2026.