All the unmentioned details from Apple's press conference are here: A19 Pro, heat dissipation, and self-developed baseband
Early in the morning on September 10th, Apple released four new models in the iPhone 17 series and simultaneously introduced the new-generation A19 chip.
Before the press conference, external expectations for the A19 mainly focused on quantitative indicators such as performance improvement and an increase in the number of transistors, which was generally considered a routine upgrade.
Kaiann Drance (Image source: Apple's autumn new product launch event)
However, during the one-hour and twelve-minute press conference, although the A19 and A19 Pro made three appearances, Kaiann Drance, the vice president of global marketing for iPhone, and Tim Millet, the vice president of wireless software technology, who were responsible for the introduction, provided very little information about the specific improvements of the A19 in terms of CPU and did not disclose more details.
Tim Millet (Image source: Apple's autumn new product launch event)
After the press conference, we had the opportunity to have a face-to-face conversation with Kaiann Drance and Tim Millet, who had just stepped off the stage, to delve into this chip and thus learned more behind-the-scenes stories about the A19.
The "Secret" Hidden in the GPU
Apple's description of the A19 process (Image source: Apple China official website)
After the latest product page update, Apple clearly marked that the A19 chip uses the third-generation 3nm process. This information perfectly matches the pre-press conference rumor that "it will be equipped with TSMC's N3P process", confirming the previous external speculation.
Regarding the CPU performance improvement, Kaiann disclosed for the first time after the press conference that compared with last year's A18 Pro, the processor performance of this year's A19 Pro has increased by about 10%. However, this upgrade amplitude basically did not exceed the pre-press conference expectations of the outside world for this chip. That's why Apple provided very little information about the CPU performance improvement during the press conference. The iteration from the N3E process to the N3P process itself is relatively stable. For Apple, which is good at telling "revolutionary" technology stories, a 10% performance improvement is indeed not enough to be a highlight emphasized at the press conference.
In fact, the real core upgrade of the A19 Pro is hidden in the lower right corner of the following picture - the GPU neural network accelerator.
The GPU neural network accelerator appears for the first time on the A19 Pro (Image source: Apple's autumn new product launch event)
Limited information about this function was disclosed at the press conference, but it has become the key improvement that the outside world has interpreted the most: Apple has added a neural network accelerator to each core of the A19 Pro GPU. According to official data, its peak GPU computing power has increased to three times that of the A18 Pro.
This design is similar to NVIDIA's Tensor Core and is different from the "Neural Engine" that Apple has long used, which is suitable for low-power scenarios.
When asked why they designed the GPU accelerator, Tim Millet said that "efficiency improvement" was their core concern.
Tim Millet gave an example to the author. Originally, on A-series processors without a GPU neural network accelerator, if a complex local inference operation needed to be processed, the GPU would send this task to the Neural Engine through unified memory for processing. After the processing was completed, the Neural Engine would then send the data results back to the GPU.
Although the unified memory has made this process fast enough, Apple found that even with a short data transmission path, there was still a waste of resources due to data exchange. The design of the built-in GPU accelerator allows tasks that could originally be processed on the GPU side to be completed directly inside the GPU, reducing the data exchange link, even if it only improves the efficiency by one-thousandth of a second.
However, this design also has an obvious contradiction: although the GPU accelerator has stronger performance and higher efficiency, it often comes with higher energy consumption when dealing with high-performance model scenarios.
At the same time, limited by the form factor of mobile phones, the on-device inference and generative capabilities of the A19 Pro are still limited by hardware conditions such as memory bandwidth, available memory, and chip size. If it is only for a "little bit of efficiency improvement", such a design will inevitably make people feel "overdone".
In response to this question, Tim Millet added Apple's second design consideration: In current mobile phone usage scenarios, pure artificial intelligence inference or generative tasks (such as convolutional neural networks CNN, long short-term memory networks LSTM, Transformer models, etc.) are more suitable to be completed in the NPU (Neural Processing Unit). But in reality, most mobile phone tasks are "hybrid workflows" that do not rely on a single core to complete tasks but require the coordinated operation of the CPU, GPU, NPU, and unified memory. The original design intention of the GPU neural network accelerator is to solve complex task scenarios in hybrid workflows. For example, the MetalFX super-resolution technology will be significantly improved in this upgrade.
Combining the latest information provided by Apple, the neural network accelerator of the GPU can work in tandem with the new 16-core neural engine. Coupled with the update of the dynamic cache, the computing performance of FP16 has doubled, a new unified image compression technology has been built, and with the continuous improvement of dynamic memory bandwidth, with the support of the A19 Pro, the iPhone 17 Pro, Pro Max, and the brand-new iPhone Air this year can all better handle some AI tasks that have high requirements for the GPU, such as running large language models locally and playing games with extremely high graphics processing requirements, achieving hardware-accelerated ray tracing and higher frame rates.
Then, let's make a bold assumption. If the results of Apple's "experiment" in this step are verified to be positive, the M5 chip that may be unveiled in Q4 this year may also adopt a similar design. This means that the AI computing power of the M5's GPU will be pushed to an unprecedented height.
The M3 Ultra released by Apple in March this year has increased the bandwidth of the unified memory to 819GB/s. If the future M4 Ultra or M5 Max can be paired with faster LPDDR 6 memory, pushing the bandwidth above 900GB/s, combined with the experimental results of Apple's GPU neural network accelerator on the A19 Pro, the on-device computing power of future Macs cannot be underestimated.
Replacing Titanium with Aluminum, a Gamble by Apple
The design of the iPhone 17 Pro has become a hot topic, mainly because Apple has replaced the titanium middle frame with an aluminum alloy, bringing the Pro model to the same level as the standard version.
Behind this "downgrade" move is to make way for a new active cooling system - the thermal conductivity of aluminum is 20 times that of titanium, making it more suitable for heat dissipation requirements.
Although active cooling is already common in the Android camp, the VC vapor chamber is used on an iPhone for the first time. Apple has encapsulated it on the back of the motherboard, directly connected to the middle frame. The deionized water inside efficiently dissipates heat through a vaporization/liquefaction cycle, specifically serving the A19 Pro chip.
The VC vapor chamber used by Apple in the iPhone 17 Pro series this year (Image source: Apple's autumn new product launch event)
Before this year, Apple used more traditional graphite as the chip's cooling system. The upgrade to the VC vapor chamber this year shows that Apple has higher requirements for the continuous high-performance output of the A19 Pro.
Kaiann Drance introduced that tests showed that in an extreme continuous working environment, the A19 Pro equipped with a VC vapor chamber had a 40% longer continuous working time compared to the A18 Pro without a VC vapor chamber.
This means that starting this year, Apple has truly taken seriously the heat dissipation challenges brought by three high-load scenarios in daily applications:
● Improved gaming performance: The GPU neural network accelerator of the A19 Pro has enhanced the rendering ability, but it has also increased the heat dissipation pressure.
● Upgraded image processing: The new lens system and the imaging positioning of the Pro series have led to increased heat generation during shooting.
● Increased AI tasks: The device undertakes more inference and generative AI tasks under the existing hardware conditions.
The display driver unit introduced on the A19 chip (Image source: Apple's autumn new product launch event)
In addition to heat dissipation, Apple has also placed a big bet on the standard version of the iPhone 17. The A19 chip integrates a display driver unit for the first time, finally bringing a 120Hz high-refresh screen to the standard version.
When asked about the reason for such an adjustment to the standard version, Kaiann Drance said that this year was the best time. Having the display unit on the A19, including the change in the screen material this year, all had the same consideration.
But such an answer is obviously unconvincing. Data on the first three weeks of the iPhone 16 series released by Counterpoint in October last year showed that the iPhone 16 series had a strong start in the Chinese market. Compared with the iPhone 15 in 2023, the sales volume in the first three weeks after its launch increased by 20%. Among them, the high-end models Pro and Pro Max of the iPhone 16 sold particularly well, while the growth rates of the iPhone 16 and 16 Plus were far lower than those of the two 16 Pro models.
Similar data appeared in June last year. Counterpoint's report showed that before the 618 shopping festival, thanks to the promotional activities of the iPhone 16 series, especially the increased sales of the iPhone 16 Pro and iPhone 16 Pro Max, Apple achieved a year-on-year increase in sales volume in May.
All signs indicate that the standard models of the iPhone have been at a disadvantage in the market competition cycle for at least the past two years, which has become one of the driving forces for Apple to accelerate the adjustment of the iPhone's design and product lineup.
This strategic adjustment has been immediately effective. In the domestic pre-sale performance of the iPhone 17, the standard version has surpassed the Pro for the first time to become the most popular model, and the addition of the high-refresh screen has played a significant role.
Through the design adjustment of "replacing titanium with aluminum" and the key upgrade of the standard version, Apple has demonstrated its flexibility in product strategy, ensuring the performance release of the Pro series and successfully enhancing the market competitiveness of the standard version.
Is the iPhone Air a Better Product?
When asked by Tim Millet whether the "iPhone Air is theoretically a better product", Tim Millet smiled knowingly. This phone indeed embodies many "firsts" for Apple: it is the first iPhone product designed with a new concept of thinness and lightness, the N1 and C1X chips it is equipped with are making their debut, and it is the first iPhone model to completely eliminate the physical SIM card.
When talking about the chips, Tim Millet said, The iPhone Air is equipped with the A19 Pro chip, ensuring flagship-level performance. The addition of the N1 and C1X chips reflects Apple's core design philosophy: self-developed chips are for achieving better functions and product innovation, rather than simply pursuing parameters.
This makes the Air Apple's biggest market test field this year, testing two key issues at the same time: the market demand for thin and light mobile phones and the stability of the coordinated operation of multiple self-developed chips.
When asked why only the Air is equipped with the C1X baseband chip, Tim Millet explained to the author. The C1X was designed for the iPhone Air. If you want to achieve a thinner and lighter design, you must also consider energy efficiency, and the C1X has a great advantage in this regard.
Data released at Apple's press conference showed that the performance of the C1X is twice that of the C1, faster than the Snapdragon X16 running in the iPhone 15 Pro and iPhone 16 Pro Max, and its power consumption is reduced by 30%.
The network differences of iPhones this year (Image source: Apple's US official website)
However, if you look at the official parameter table of the US version of the iPhone, you will find that the iPhone Air equipped with the C1X currently does not support 5G millimeter-wave technology. If you choose the iPhone Air, you can only let its network operate in the low-frequency band below 6GHz, which means that for US users, choosing the Air means giving up faster upload and download network speeds.
This may also be the key factor why Apple has not fully replaced Qualcomm's baseband chips with its own. However, this does not affect Apple's determination to continue promoting self-developed chips.
The emergence of the C1 and C1X has indeed put Qualcomm in a difficult position. Apple and Qualcomm's licensing agreement will expire in 2027. Some analysts predict that the proportion of Qualcomm's baseband chips used in next year's iPhones will drop to 20%, and by 2027, Qualcomm is expected to completely exit Apple's supply chain.
Qualcomm CEO Cristiano Amon also said in an interview this year that according to the existing contract, if they cannot obtain