Apple Announced the M6 First. It Gave Away Most of the A20 Pro.

An order Apple had kept for a decade

Apple's A-series and M-series chips share most of their architecture within a generation: the CPU cores, the GPU design, the Neural Engine and the media block are largely the same silicon, scaled up for the Mac. In practice, the iPhone chip announced in September has functioned as a partial preview of the Mac chip that arrives months later.

Apple inverted that this year. The M6 was revealed on 25 August in the new Mac mini, and the A20 Pro is expected on Wednesday in the iPhone 18 Pro. As MacRumors set out this weekend, a shipping Mac chip now documents much of what the iPhone chip will do. Reports place the A20 Pro at up to 18 percent faster and up to 30 percent more power efficient than the A19 Pro, broadly in line with what the M6 delivers over the M5.

The 2 nm claim was spent two weeks early

The M6 is Apple's first chip built on TSMC's N2 process. The A20 is almost certainly the second.

That matters commercially more than technically. Apple has historically introduced each new manufacturing node on the A-series, because iPhone volumes are what make a new node's yields economic: hundreds of millions of small dies absorb defects far better than a few million large ones. Debuting N2 in an $899 Mac mini is a departure from that logic, and it suggests TSMC's yields were healthy enough that Apple did not need iPhone volume to derisk them.

The side effect is that the most quotable line of Wednesday's keynote is no longer available. "Apple's first 2 nanometre chip" was announced a fortnight ago, to far less attention.

A second Neural Engine, and where it may not fit

The iPhone's Neural Engine has stayed at 16 cores and roughly 35 trillion operations per second since the A17 Pro, three generations without a headline change. The M6 introduces a dual 16-core Neural Engine that system frameworks can address simultaneously, which Apple rates at up to twice the peak compute of the previous generation.

That change lines up with the reported move to WMCM packaging on the A20, which places memory on the same wafer as the CPU, GPU and Neural Engine instead of stacking it on top. Shortening the physical distance between memory and the units that read it is one of the few ways left to make a mobile chip faster without spending more power.

Whether a second Neural Engine fits inside a phone is a separate question from whether it works. Die area and heat are the constraints, and a dual engine may end up reserved for the A20 Pro rather than shipping across the whole A20 family.

Memory bandwidth is the ceiling nobody puts on a slide

This is the number that will decide how the iPhone 18 Pro handles on-device AI, and it is the least discussed.

Running a language model locally is mostly an exercise in moving weights from memory to the processor. Every token generated requires reading a large fraction of the model, so generation speed tracks memory bandwidth far more closely than it tracks raw compute. A chip can have an enormous Neural Engine and still produce text slowly if the memory cannot feed it.

The M6 provides up to 170 GB/s, which Apple describes as a 10 percent increase over the M5's 153.6 GB/s. The A19 Pro is rated at 76.8 GB/s. A comparable uplift would put the A20 Pro near 85 GB/s, around half the bandwidth of a Mac mini, which is a reasonable way to understand why phones run small models and Macs run larger ones.

One caution on reading that figure too confidently: bandwidth alone does not identify the memory generation. The same 170 GB/s can be reached with fast LPDDR5X on a wide bus or with LPDDR6 on a narrower one. What can be said is that the increase is modest, and that any real gain in on-device model performance is more likely to come from packaging than from the memory itself.

Graphics: the architecture travels, the headline does not

The M6 adds 50 percent higher geometry rates, a revised shader core architecture, updated Dynamic Caching and hardware-accelerated ray tracing. Changes of that kind have historically appeared on both chip families within the same generation, so the iPhone should inherit them.

The headline number is another matter. Apple said the M5 delivered over four times the peak GPU compute for AI compared with the M4, thanks to Neural Accelerators inside each GPU core; the A19 Pro recorded a similar four-fold gain last September for the same reason. The M6 is rated at nearly 30 percent over the M5. Second-generation Neural Accelerators look like a refinement, which means the A20 Pro is unlikely to repeat last year's multiplier.

AV1 encode is still absent, and that is a decision

The M6 media engine handles H.264, HEVC, ProRes and ProRes RAW, with AV1 decode but no AV1 encode, the same as the M5. AV1 is the royalty-free codec used by YouTube and most streaming services, and hardware encoding is what makes it usable for uploads and video calls rather than only for playback. Several Android flagships have shipped it.

Leaving it out of a desktop chip on a new node makes its arrival in a phone chip this year very unlikely. Video uploaded from an iPhone 18 Pro will still be H.264 or HEVC.

Our earlier analysis of the A20 Pro's 2 nm numbers covered what the process claim does and does not mean. The M6 has now settled much of the rest. What Wednesday will genuinely add is the packaging, the thermal behaviour, and whether the dual Neural Engine crosses over to the phone.