In-depth丨Dimensity 9400 released, mobile phone processor uses PC-level CPU architecture
Focus: artificial intelligence, chips and other industries
Welcome all guests to follow and forward
On October 9, MediaTek officially released its flagship 5G intelligent AI chip - Dimensity 9400.
The chip uses TSMC's second-generation 3nm process technology and is equipped with a second-generation full-large core architecture CPU.
The CPU of this generation of flagship chip has upgraded one of the four super-performance cores of Dimensity 9300 to Arm’s latest Cortex-X925 core (main frequency up to 3.62GHz).
At the same time, it is also equipped with 3 Cortex-X4 super cores (main frequency is 3.3GHz) and 4 Cortex-A720 large cores (main frequency is 2.4GHz).
These configurations enable the Dimensity 9400 to improve single-core performance by 35% and multi-core performance by 28% compared to the Dimensity 9300.
MediaTek pursues the dual advantages of low energy consumption and high performance, rather than simply pursuing the ultimate IP frequency.
This goal is achieved through the new CPU all-big core architecture.
Among them, the Cortex-X925 core is given the reputation of "super-large core" because of its main frequency of up to 3.62GHz and PC-level IPC performance.
As Arm's latest performance core, Cortex-X925 is designed for 3nm process and adopts PC-level architecture design, providing more powerful performance support for the processor.
In addition, Cortex-X925 also introduces vector pipeline technology, which achieves a 50% improvement in Integer8 computing performance (TOPS).
However, despite the excellent performance of the Cortex-X925 in terms of main frequency, it still falls short of the main frequency of the CPU super core of Qualcomm's upcoming Snapdragon 8 Gen 4.
This reflects that in the current context of the slowdown of Moore's Law, IC performance has shown a trend of high redundancy or even excess.
Therefore, how to find the best balance between performance and power consumption to achieve an excellent design experience has become an important challenge facing chip designers.
The all-core CPU architecture design and careful consideration of IP selection adopted by MediaTek in Dimensity 9400 are a powerful response to this challenge.
Compared with its predecessor, Dimensity 9300, Dimensity 9400 has achieved significant performance improvements, with a 35% increase in single-core performance and a 28% increase in multi-core performance.
In the AnTuTu benchmark test, the Dimensity 9400 has exceeded 3 million points in a laboratory environment, which further confirms its powerful performance.
GPU and NPU improvements further improve AI performance
In terms of GPU, the Dimensity 9400 is equipped with a new top-level 12-core GPU Immortalis-G925. Its peak performance has been significantly improved compared to the previous generation, reaching an increase of 41%. At the same time, power consumption has been reduced by 44%, and ray tracing performance has also been improved by 40%.
In addition, Dimensity 9400 also supports the PC-level Dimensity OMM light chasing engine and Star Speed Engine super-resolution technology. Both technologies represent the most advanced level in the industry and support the industry's most advanced interpolation technology and super-resolution technology.
From these data, we can see that Dimensity 9400 has achieved a performance leap in both CPU and GPU, while power consumption control has also reached an extremely high level. This is undoubtedly a wonderful application of the principles of physics in technological innovation.
In terms of NPU, Dimensity 9400 integrates the eighth-generation AI processor NPU 890. MediaTek summarizes its outstanding capabilities with three words: [smart, strong, and dense].
The addition of NPU 890 has enabled Dimensity 9400 to achieve significant improvements in AI performance and energy efficiency.
In the ETHZAIBenchmark v6.0 benchmark test, Dimensity 9400 achieved an excellent score of 6773 points, ranking first in the AI performance of mobile phone chips.
With the leap-forward improvement in the performance of CGN (i.e. CPU, GPU, NPU), the gap between its and memory transmission speeds has become increasingly prominent. This phenomenon is vividly called the "memory wall" by the industry, and it has become a key factor restricting the improvement of the overall computing power of the chip.
To effectively meet this challenge, Dimensity 9400 pioneered the introduction of Samsung's LPDDR5X 10.7Gbps technology, which leads a new trend in the mobile memory field and achieves a significant 25% performance improvement compared to previous generation products.
In addition, Dimensity 9400 relies on its unique [Black Hawk] architecture design to further optimize the cache layout, which greatly shortens the distance between the cache and CGN, thereby greatly reducing the time required for data transmission.
In terms of specific capabilities, Dimensity 9400 supports end-side LoRA training and is the first to launch the end-side high-definition video generation function.
At the same time, the prompt word processing performance of its large language model has been improved by 80%, further enhancing the ability to understand the relative relationship between pictures, text, and numbers, while power consumption has been reduced by 35%.
In addition, Dimensity 9400 is also compatible with mainstream large models on the market, including Alibaba Cloud, Baichuan Intelligence, and Google.
Supplement AI application scenarios and realize cross-application connection of AI
Currently, the development of the AI industry is facing challenges, and the products of major AI star companies are still insufficient in specific application scenarios.
However, the recent Nobel Prize in Physics awarded to the [Godfather of AI] has undoubtedly injected new vitality into the field of AI.
On the smartphone, an ideal platform for AI applications, after months of software and hardware iterations, the launch of Dimensity 9400 has brought a new AI experience to users.
MediaTek is actively working with APP developers from various industries to provide a unified standard interface between AI agents, third-party applications and large models to achieve cross-application connection of AI.
The introduction of AI intelligent entity engine has pushed traditional AI applications to a higher level of intelligent entity application.
The device integrates DAE (Dimensity AgenticAIEngine), an AI intelligent engine that gives the phone the ability to actively think, plan and execute tasks.
It significantly enhances the device's autonomous learning and decision-making capabilities, and can actively adapt to and optimize user habits, thereby providing a more personalized and intelligent service experience.
Dimensity 9400 achieves efficient optimization of large models through a series of innovative technologies, including hardware-supported low-bit-width KV cache technology, the application of GQA (graph query language), and optimization strategies for memory usage.
Together, these measures improve the speed and accuracy of AI tasks.
In terms of AI edge implementation, Dimensity 9400 breaks the traditional [cloud training, edge inference] model and launches the industry's first LoRA training technology.
This innovation enables users to train models directly on local devices, achieving dual acceleration of inference and training.
This not only reduces dependence on cloud resources, but also significantly reduces data transmission delays and bandwidth consumption, thereby improving overall processing speed.
At the same time, it also helps users better control their personal privacy and effectively avoid the risk of data leakage.
Another outstanding advantage of on-device AI training is that it can deepen the mobile phone AI's understanding of users.
Compared with the traditional cloud-side training model, terminal-side training can perform continuous cumulative training based on users' daily usage data.
This means that mobile AI will be able to more accurately capture user preferences and habits, continuously optimizing the service experience over time.
AI revolutionizes the mobile phone industry and opens a new round of innovation cycle
Although Apple launched the iPhone 16 series in September 2024, its sales after its launch fell far short of market expectations due to its lack of innovation momentum.
In sharp contrast, new technological innovations represented by artificial intelligence are leading the smartphone industry into a [new round of innovation cycle].
In this context, 2024 has been recognized by the industry as the first year of AI mobile phone popularization. According to the forecast of well-known research institutions, the global shipment of new generation AI mobile phones will reach 170 million units in that year, accounting for about 15% of the total shipment of smartphones.
What is even more remarkable is that the compound growth rate of AI mobile phones is expected to reach 73.47% by 2027.
By 2025, the ability of flagship mobile chips to support edge AI will become one of the key factors determining their competitive advantage.
Since 2023, many mobile phone brands have begun to integrate self-developed AI big models into their mobile phone products.
Compared with cloud-based AI, edge AI has higher data security, faster response speed, and smoother and more stable operation performance because it is deployed locally.
However, the core challenge is that due to the limitations of mobile phone computing power, the parameter scale of most current large models on the edge is relatively small, which in turn limits the further expansion of the AI experience and makes it necessary to rely on cloud AI support when processing more complex AI tasks.
Therefore, whoever can be the first to achieve technological breakthroughs in the edge AI field and subvert the traditional AI experience in the future will be able to occupy an advantageous position in the upcoming market competition.
In this process, a crucial and fundamental link is to have advanced chips with higher AI computing capabilities and the ability to support larger-scale parameter-side AI.
SoC defines AI mobile phones, 3nm brings new opportunities
With the release of the first half sales data, the recovery trend of the smartphone market has been further confirmed.
In this process, [AI mobile phone], as a frequently mentioned keyword, has become an important force driving market recovery.
The big model entry port marks a new stage in the development of AI big models. Smartphones, as the most important mobile terminal in the past decade, have been given a dual mission:
It is necessary to promote the commercialization of big models and use big models to create new growth points for themselves.
However, compared with the enthusiasm of mobile phone brands and institutions, consumers still lack a clear understanding of the concept of AI mobile phones, and the user experience also has many shortcomings.
This is mainly due to the fact that the hardware core of AI mobile phones - SoC (system-on-chip) has not yet reached its ideal state.
Authoritative research organizations IDC and Counterpoint have both given clear definitions of AI phones, and unanimously regard mobile phone SoC as an important criterion for judging AI phones.
IDC believes that AI phones should be equipped with mobile chips that meet AI computing power requirements and integrate deep learning AI functions. Their NPU (neural network processor) computing power must be greater than 30 TOPS.
Counterpoint emphasizes that AI mobile phones must have integrated or independent neural network computing units (such as APU, NPU, TPU) to ensure powerful AI computing power.
Judging from the AI mobile phones on the market according to the above standards, there are not many products that truly meet the requirements.
At present, only a few SoCs such as Dimensity 9300, Apple A17 Pro, and Qualcomm Snapdragon 8Gen3 have reached the [passing line] standard in computing power, and the number of AI mobile phones equipped with these chips is relatively limited.
Compared with the first-generation 3nm process used by Apple A17 Pro, the second-generation 3nm process used by Dimensity 9400 is more technically mature.
At the same time, compared with the 4nm process used by Dimensity 9300, Dimensity 9400 also has more room for overall design.
Ending:
The mobile phone SoC chip market has become more stable. In the Android camp, apart from the unique position of iPhone, its open chip ecosystem mainly revolves around the two major brands of MediaTek and Qualcomm.
The birth of AI phones has closely linked competition in the smartphone market with mobile phone SoCs, and has also shifted the pressure of innovation from mobile phone brands to SoC manufacturers.
Some references: Wall Street Journal: "Dimensity 9400: Deconstruction of the design ideas of the top SoC chip", Computer Daily: "Why is Dimensity 9400 the AI chip of the next era?", Lei Technology: "Dimensity 9400 first test: stronger full-core, stable flagship U", Core Things: "Overtaking Apple in three years, from energy efficiency to edge AI, MediaTek sets a new benchmark for high-end chips", Silicon Star Pro: "MediaTek fired the first shot of mobile phone chips [volume AI]", One Observation: "Dimensity 9400: The core of the trend, the way to win in the future", Yuanchuan Research Institute: "Dimensity 9400 is released: the competition of flagship SoC is upgraded", Huxiu APP: "Who is the driving force behind high-end flagship mobile phones?"
The articles and pictures published on this public account are from the Internet and are for communication purposes only. If there is any infringement, please contact us and reply. We will deal with it within 24 hours after receiving the information.
Recommended reading:
京公网安备 11010802033920号