A20 Pro's memory bandwidth gets a 50% boost, Apple says, to help handle larger AI models
Apple gave the iPhone 18 Pro a 32-core neural engine and FP8 support, but the figure that decides whether your model runs at usable speed is bandwidth, and it arrives as a percentage with no gigabytes per second attached.
The Engineer · Build desk
What happened
- Apple's A20 Pro carries 32 neural engine cores split across two engines, adds support for 8-bit floating-point operations, and raises memory bandwidth 50% over the previous generation.
- It is Apple's first 2nm chip in an iPhone, pairing a six-core CPU whose two performance cores are 20% faster with a seven-core GPU Apple claims is 40% faster and carries new neural accelerators.
- Apple says the phone runs its most advanced on-device models yet, with the onscreen-aware Siri arriving in English beta alongside iOS 27 this month and five more languages in October.
- The iPhone 18 Pro starts at $1,199 and the Pro Max at $1,299, each $100 above last year's model, with sales opening on September 18.
Why it matters
- What it limits constraint Nobody can size a model against this spec sheet: with percentages instead of gigabytes per second and no confirmed DRAM capacity, the fit question only gets answered on hardware.
- Who has to decide decision Which FP8 claim an app actually collects depends on whether its matmul is dispatched to the GPU's neural accelerators or to the neural engine, so the dispatch target becomes a shipping decision rather than an implementation detail.
- What's now possible capability Minutes-long generation loops, not benchmark bursts, are the workload this packaging is tuned for, which makes sustained throughput the honest measure of what an app can attempt.
- What it costs cost Any feature that needs the full envelope runs only on the three devices carrying the A20 Pro, so the installed base for it is the priciest slice of the line and grows one release at a time.
Next story · Product · 12h ago
Apple opens iPhone Duo preorders Friday at 5 a.m. PT, with U.S. pricing starting at $1,999Confirmed9 publishers
Bandwidth decides whether an on-device model is usable, and Apple gave it as a ratio: 50% more than the previous generation, which Tom's Hardware reports Apple calls the widest memory interface in an iPhone [4]. No gigabytes-per-second number is attached to either end of that comparison in the launch reporting [22]. That matters because a decoder-only model generating one token at a time is memory bound. Every active weight is read out of DRAM per token, so throughput tracks bandwidth rather than multiply-accumulate capacity. A 50% bandwidth increase is close to a 50% lift in the tokens-per-second ceiling for that workload, and that is the load-bearing version of Apple's claim that the chip handles larger AI models [3].
Capacity is the other half, and it is absent [22]. A pre-announcement rumour quoted in a reader reply on the Tom's Hardware story put the configuration at a 96-bit 12GB LPDDR5X-9600 interface with 36MiB of SLC cache; Apple confirmed none of it [21]. Capacity determines whether an 8-bit quantised model loads at all. Bandwidth only determines how fast it runs once it has.
FP8 arrives twice, from two different blocks. The Economic Times reports the 32-core neural engine supports 8-bit floating point [1]. Tom's Hardware reports the seven-core GPU's new neural accelerators deliver twice the FP8 compute [17]. The doubling is attached to the GPU. Worth separating FP8 from int8 as well: FP8 buys dynamic range at the same byte count, which helps with activation outliers, but it does not make weights smaller than int8 already made them. The win is accuracy at eight bits, not a bigger model in the same DRAM. The 32 cores are also two engines rather than one [2], and whether a single graph can span both is not stated [26].
The packaging is where the actual engineering shows. Memory comes out of the SoC's thermal path so the die attaches directly to the vapour chamber [8], the chamber has three times the surface area of the iPhone 17 Pro's [9], and a nano-twin copper material pulls heat away from the display [10]. Apple claims up to 40% sustained performance over the 17 Pro and 2x over the 16 Pro [24]. Divide those: 2.0 / 1.4 puts the 17 Pro at about 1.43x the 16 Pro on the same measure [25], so roughly half the two-generation sustained gain predates this chip. The silicon story this year is partly a heatsink.
Sustained is also the figure that transfers to inference. A token loop runs for seconds or minutes, well past any burst window, which is why the 20% CPU improvement [15] describes work you would not put a matmul on.
The battery numbers are not inference numbers. The 36 and 45 hours are video playback [6]; Apple's own usage-model test gives 24 and 30 hours [7], a factor of 1.5 between the two on the Pro [20]. Video playback runs largely on fixed-function decode. Neither test tells you what a half-hour summarisation session costs.
The bandwidth figure reaches your app if the model is memory bound and fits in package DRAM. The sustained figure reaches it if the workload outlasts the burst window. The FP8 doubling reaches it only when the matmul lands on the GPU accelerators rather than the neural engine, and nothing in the launch material describes the API surface that decides where it lands.
What to watch
- Whether Apple's published iPhone 18 Pro tech specs give an absolute memory bandwidth figure or DRAM capacity, which would let developers size a model before shipping.
- A Core ML or Foundation Models release note stating which unit an 8-bit floating-point matmul dispatches to, and whether both neural engines are addressable as one device.
- Teardowns after the September 18 on-sale date, which would confirm or kill the 96-bit LPDDR5X-9600 configuration described in pre-announcement rumours.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence62
- Adoption45
- Hype gap+34
- Incentives71
- Confidence66
Perspective Coverage
7 publishers- Builder
- Builder 49%
- Operator
- Operator 27%
- Investor
- Investor 24%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The A20 Pro's new neural engine is designed for on-device AI workloads, with a 32-core design and support for advanced 8-bit floating-point operations.
- [2]
The A20 Pro has two neural engines with a combined 32 total cores.
- [3]
Apple increased memory bandwidth by 50%, which it says helps the device handle larger AI models and demanding applications more efficiently.
- [4]
There is a 50% increase in memory bandwidth on the chip, which Apple says is the widest memory interface in an iPhone.
- [5]
The A20 Pro is Apple's first 2-nanometer chip in an iPhone; Apple's first-ever 2nm chip is the M6, announced in August, debuting in the Mac Mini later this month.
- [6]
The iPhone 18 Pro offers up to 36 hours of video playback and the Pro Max up to 45 hours, the longest battery life ever offered on an iPhone.
- [7]
Using a proprietary test based on data from how people use their phones, Apple claims 24 hours per charge on the iPhone 18 Pro and 30 hours on the Pro Max.
- [8]
Apple changed the packaging, placing the silicon dies so that memory is removed from the thermal path of the SoC, allowing the silicon to attach directly to the vapor chamber.
- [9]
The vapor chamber has three times larger surface area than the iPhone 17 Pro's and includes more graphite and copper along with 80% recycled stainless steel.
- [10]
Apple introduced new thermal materials including a nano-twin copper material designed to move heat away from the display more effectively.
- [11]
The iPhone 18 Pro starts at $1,199 and the Pro Max at $1,299, with availability on September 18.
- [12]
Both models are $100 more expensive than last year's devices.
- [13]
Apple says the iPhone 18 Pro can run its most advanced on-device models yet, with access to Apple's latest Foundation models.
- [14]
With iOS 27 this month, the new Siri with onscreen context analysis becomes available in English in beta, with French, Japanese, Korean, Portuguese and Spanish planned for October.
- [15]
The A20 Pro has a 6-core CPU with two super-cores that are 20% faster than last generation, plus four efficiency cores with neural accelerators; Apple calls it a desktop-class processor and the fastest CPU in a smartphone.
- [16]
The A20 Pro will power the foldable iPhone Duo along with the iPhone 18 Pro and Pro Max.
- [17]
The 7-core GPU has a 40% boost generation over generation with increased bandwidth, along with new neural accelerators that Apple says allow twice as fast FP8 compute.
- [18]
Apple's new in-house C2 modem replaces Qualcomm, and Apple claims meaningfully faster uploads than the C1X while consuming 15 percent less energy, adding mmWave support in the United States.
- [19]
The iPhone 17 Pro Max is powered by the A19 Pro, which has a six-core CPU and a six-core GPU.
- [20]
On the iPhone 18 Pro, the video playback figure is 1.5 times the figure from Apple's proprietary usage-based test.
- [21]
A pre-announcement rumour article quoted in a reader reply on the Tom's Hardware story claimed performance cores up to 4.9 GHz, efficiency cores up to 2.6 GHz, 36 MiB of SLC cache, and a 96-bit 12 GB LPDDR5X-9600 memory configuration.
- [22]
The launch reporting states the A20 Pro memory improvement only as a 50% bandwidth increase, with no gigabytes-per-second figure and no confirmed DRAM capacity.
- [23]
Going from the A19 Pro's six GPU cores to the A20 Pro's seven is a 16.7% increase in core count, against a claimed 40% generational graphics gain.
- [24]
Apple claims the design allows up to 40% sustained performance over the iPhone 17 Pro and 2x sustained performance over the iPhone 16 Pro.
- [25]
Apple's two claims imply the iPhone 17 Pro already ran at about 1.43x the iPhone 16 Pro's sustained performance.
- [26]
The launch reporting does not state whether a single model can address all 32 cores across both of the A20 Pro's neural engines.
Sources
7 independent publishers whose own reporting we read for this story.
- archive.thedeepview.comiPhone Duo unfolds Apple's vision of the AI future
1 article · September 10, 2026
- dev.to💻 What Apple’s September 2026 Event Means for Developers
1 article · September 10, 2026
- economictimes.indiatimes.comThe foldable iPhone as Apple turns the page: John Ternus faces his first big test as CEO - The Economic Times
9 articles · September 10, 2026
- mezha.netApple може представити складаний iPhone: новинка стане випробуванням для Джона Тернуса
4 articles · September 10, 2026
- techspot.comThe foldable iPhone Duo is expected to headline Apple's event today – watch it here at 1pm ET / 10am PT | TechSpot
6 articles · September 12, 2026
- theneuron.aiEverything AI Apple Announced at Its September 9 Event
3 articles · September 9, 2026
- tomshardware.comApple A20 Pro powers iPhone Duo, 18 Pro — the company's first 2-nanometer smartphone chip
1 article · September 9, 2026
Topics and entities
Topics
- Mobile Silicon and Modems
- On-Device AI
- Foldable Smartphones
- Memory Bandwidth