Apple has announced two new chips, M6 and M5 Ultra, with the official headline naming “performance and AI compute” outright. The M6 lands in the Mac mini and the M5 Ultra in the Mac Studio; both are now open for pre-order, with deliveries starting September 22.
The M6 is Apple's first 2-nanometer chip, and one of the first mass-produced products on TSMC's 2nm process. It packs a 12-core CPU (2 hyper cores, 4 performance cores, 6 efficiency cores), a 12-core GPU, dual 16-core Neural Engines, up to 32GB of unified memory, and 170GB/s of memory bandwidth. Apple says it delivers the fastest single-thread performance of any chip in the world, with multi-thread performance 1.2x that of the M5. GPU AI compute is up roughly 30% over the M5 and more than 8x over the M1, while Neural Engine performance is double the previous generation.
M5 Ultra takes the brute-force route: a new generation of UltraFusion links two dual-die M5 Max chips together, making it the first four-chip architecture in the M series.
| M6 | M5 Ultra | |
|---|---|---|
| CPU | 12-core | Up to 36-core |
| GPU | 12-core | Up to 80-core |
| Neural Engine | Dual 16-core | 32-core |
| Unified memory | Up to 32GB | Up to 512GB |
| Memory bandwidth | 170GB/s | 1.2TB/s |
This time, the M5 Ultra's GPU also gains Neural Accelerators, with AI compute up 4.5x over the M3 Ultra and more than 6x over the M1 Ultra; its 1.2TB/s of bandwidth is 50% higher than the M3 Ultra. In Apple's own words: “M5 Ultra features a massive GPU, now with Neural Accelerators, and more unified memory bandwidth, pushing the boundaries of what a desktop can do.”
What Fits Inside the 512GB Ceiling
Unified memory matters most visibly in local inference: the CPU and GPU share the same memory pool, so model weights never need to shuttle back and forth between VRAM and system RAM. For discrete-GPU setups, VRAM capacity is a hard ceiling; for the Mac Studio, that ceiling is now 512GB.
A rough back-of-envelope calculation: for the open-source MoE models in the 600-700 billion parameter range, 4-bit quantized weights come in around 350GB to 400GB. Add KV cache and runtime overhead, and 512GB is just enough to fit the entire model on one machine. Reaching that same amount of usable memory on a comparable data-center setup would require linking multiple GPUs together — an entirely different order of cost and power draw.
Bandwidth is still the weak point. 1.2TB/s is a big number for a desktop machine, but it still falls well short of data-center HBM setups, and token generation speed will visibly suffer on long-context inference. What it buys instead is something else: a machine you plug into the wall, running a local model, with data that never leaves the building.
The 2nm Step
With M6, Apple gets first dibs on 2nm, staying out ahead of the pack as usual. The process gains show up in efficiency rather than peak clock speed — six of the 12 cores are efficiency cores, and bandwidth sits at 170GB/s, only 10% above the M5. M6's design clearly prioritizes keeping power draw in check, with most of the AI compute gains coming from a reworked GPU and Neural Engine rather than raw core count.
Set against the rest of the industry this same week, the direction lines up. Data centers are racing to squeeze more inference out of every kilowatt; desktops are racing to fit bigger models onto a single machine. Inference cost is being pushed down from both ends at once — just through different means.
As for the M6 in the Mac mini, capped at 32GB of memory, the scale of model it can run locally is limited — this chip's target buyer clearly isn't chasing that particular line.
Sources: Apple Newsroom official announcement, MacRumors, CocoLoop, Phoronix, 9to5Mac; core counts, memory capacity/bandwidth, and multiplier figures were cross-checked against the official announcement, and the quantized model size is a rough estimate based on public parameters.