On September 9, 2026 (US time), Apple announced the A20 Pro, the chip that powers the iPhone 18 Pro, iPhone 18 Pro Max, and the newly introduced iPhone Duo. Built on a 2nm process, it delivers up to 20% faster CPU performance and up to 40% faster GPU performance than the A19 Pro, according to Apple. Memory bandwidth is up 50%, and placing the memory beside the die lets heat escape the chip more easily. Beyond adding compute resources, the changes from the previous generation show up in how data reaches those processing units and in how performance is sustained under prolonged load.
What has increased since the A19 Pro
The A20 Pro has a 6-core CPU, a 7-core GPU, and a Neural Engine with a combined 32 cores. Set against the A19 Pro, the CPU core count is unchanged, while GPU and AI compute resources have grown.
| Item | A19 Pro | A20 Pro | Performance change from A19 Pro, as shown by Apple |
|---|---|---|---|
| CPU | 6 cores | 6 cores, with integrated neural accelerators | Up to 20% faster |
| GPU | 6 cores | 7 cores | Up to 40% faster |
| Neural Engine | 16 cores | Two sets of 16 cores, 32 in total | Twice the compute capability |
Sources: A19 Pro announcement, A20 Pro announcement, announcement of the iPhone Duo, which uses the same A20 Pro. The performance changes are figures Apple published on September 9, 2026, not independent measurements. The workloads and durations used for the CPU and GPU tests, and the absolute scores before and after, are not given in the relevant performance section.
Apple claims up to 20% faster CPU performance without adding cores. The GPU core count, meanwhile, rises by about 16.7% (calculated as (7−6)÷6×100), yet Apple's stated performance gain is up to 40%. Counting cores alone does not explain the generational gap. Improvements to each core's design and to data supply may also play a part, but their individual contributions have not been disclosed.
Memory bandwidth is also 50% wider than on the A19 Pro. That is a change in the amount of data that can be transferred per unit of time, a different metric from CPU processing speed or GPU rendering speed. When evaluating performance, it is necessary to work out whether a given workload is waiting on computation or on data transfer.
Memory moved beside the die, heat routed away from the chip
In the iPhone 18 Pro lineup, the A20 Pro die and the memory sit side by side. According to Apple, a custom package inspired by the M-series takes the memory out of the chip's heat-dissipation path, allowing the A20 Pro to connect directly to a next-generation vapor chamber.
A vapor chamber is a cooling component that spreads heat through the evaporation and condensation of a liquid inside it. Apple says the new component uses revised materials and has three times the surface area of the one in the iPhone 17 Pro generation. Both the path that carries heat away from the chip and the component that spreads it have been changed together.
In describing this thermal management, Apple says sustained performance on the iPhone 18 Pro lineup improves by up to 40% over the previous generation. The GPU's "up to 40%" and the device's sustained-performance "up to 40%" refer to different things. The former is the GPU's generational difference against the A19 Pro; the latter is the difference in performance when a device, including its cooling system, is placed under continuous load. The two cannot be added or multiplied to derive an overall performance figure.
Cooling also works differently for finishing a short task quickly and for sustaining a heavy workload. If thermal throttling can be limited, speed is easier to maintain during long gaming sessions or video editing. This change in packaging is an improvement best judged under that kind of sustained load.
How 2nm and 50% more bandwidth support on-device AI
Behind the spread of 2nm manufacturing into volume products is a generational shift in transistor structure. In its overview of N2 technology, TSMC explains that it uses first-generation nanosheet technology and began volume production in the fourth quarter of 2025. Apple's announcement, however, simply says "2nm" and does not specify whether it is N2 or N2P.
Performance differences cannot be calculated from the number in a process name alone. "2nm" is a label for a manufacturing generation, not a statement that every dimension on the chip is 2nm. Evaluating the A20 Pro means looking at the compute resources and transfer capability Apple actually added.
When running a large AI model on a device, even fast compute circuits make no progress if they are waiting for the data they need. A 50% increase in memory bandwidth is a change that eases that kind of data-supply constraint. How much faster any given task becomes, though, depends on the model's size, the computation method, and how the cache is used.
Likewise, the statement that the Neural Engine has twice the compute capability does not mean a chat app will respond twice as fast across the board. The effect varies with whether the app can take advantage of the new hardware and where in the process the time is being spent. The CPU also contains circuits that assist with AI computation, so it is hard to gauge on-device AI capability from the dedicated Neural Engine's core count alone.
Apple did not give an absolute memory bandwidth figure, memory standard, or bus width in this announcement. From the 50% increase alone, it is not possible to back-calculate and pin down a specific memory standard or bus width.
What comparing with Snapdragon and Dimensity reveals about design differences
The Snapdragon 8 Elite Gen 5 includes 18MB of dedicated GPU memory called Adreno HPM. The Dimensity 9500, meanwhile, uses a dynamic cache design that handles data for the CPU, GPU, and NPU. Competitors are also focusing on mechanisms that deliver data to processing units efficiently.
| Item | Apple A20 Pro | Snapdragon 8 Elite Gen 5 | MediaTek Dimensity 9500 |
|---|---|---|---|
| CPU configuration | 6 cores | 2 Oryon Prime cores + 6 Performance cores | 1 C1-Ultra + 3 C1-Premium + 4 C1-Pro cores |
| GPU | 7 cores | Adreno | Arm Mali-G1 Ultra MC12 |
| Dedicated AI circuitry | 32-core Neural Engine in total | Hexagon NPU | NPU 990 |
| Data-supply enhancement | 50% more memory bandwidth than the A19 Pro | 18MB HPM and Tile Memory Heap | Dynamic cache for CPU, GPU, and NPU, plus SLC management |
Sources: Apple, Qualcomm product brief, MediaTek product information. Published specifications as confirmed on September 10, 2026. Because the internal configurations of CPUs, GPUs, and AI circuits differ by company, core counts cannot be read as a cross-vendor performance metric.
This table cannot establish a performance ranking, but it does allow a comparison of how each company improves data supply. What Apple has published is wider memory bandwidth, while Qualcomm's Tile Memory Heap is a feature that optimizes GPU-side memory use and bandwidth. MediaTek describes improvements to latency and bandwidth, including management of the SLC, the shared cache on the chip.
Expanding transfer capability to external memory and reusing data on the chip to reduce transfers are approaches that can be used together. A difference in announced features does not mean one company is not using the other's method. To compare real-world performance, you would need to run the same workload that uses these mechanisms and measure both speed and power consumption.
Gaming and AI performance to check after launch
The iPhone 18 Pro and iPhone 18 Pro Max are scheduled to open for preorders on September 12 in regions including Japan, with sales starting September 18. Turning Apple's stated figures into a purchasing decision will also require independent testing of retail units.
For gaming, it will be worth looking not only at peak frame rate right after launching a game, but also at frame rate and power consumption after sustained load. Matching the graphics settings and rendering resolution of the same game and recording ambient temperature would make it easier to judge how much the new thermal path actually improves playability.
For on-device AI, the time to first response and the generation speed afterward should be measured separately, using the same model and quantization conditions. Neither whether upgrading from the A19 Pro is worthwhile nor how it stacks up against other makers' phones can be settled by the GPU's maximum figure alone. For people who use demanding games or AI workloads for long stretches, the evidence for whether this packaging change matters is speed and power draw under continued use.