Alibaba Cloud joined hands with MediaTek to go on a downward spiral in the mobile phone market

The deployment of large AI models on mobile devices is accelerating day by day.

On March 28, Alibaba Cloud and MediaTek jointly announced that Tongyi Qianwen's 1.8 billion and 4 billion parameter large models have been successfully deployed on the Dimensity 9300 mobile platform. It is reported that this is the first time that the Tongyi large model has completed chip-level software and hardware adaptation, and it also means that the exploration of Model-on-Chip has officially moved from verification to a new stage of commercialization.

Behind this, terminals with their own computing power are considered to be one of the important scenarios for the implementation of AI large models, especially mobile phones that are used very frequently. At the same time, the outside world also expects that the AI model can stir up the mobile phone market, which has been weak for a long time.

Acceleration of large models on the device side

When deploying a large AI model on a mobile terminal, the computing power of the mobile phone determines the performance of the large model.

The reporter learned from Alibaba Cloud that the 1.8 billion parameter open-source model of Tongyi Qianwen far exceeds the performance of the previous SOTA model on multiple authoritative test sets, and the inference 2048 token uses only 1.8G memory at the minimum, which is a low-cost, easy-to-deploy, and commercial-friendly small-size model. The Dimensity 9300 integrates MediaTek's 7th generation AI processor, the APU790, which delivers up to eight times the processing speed of the previous generation of AI processors.

In addition, Alibaba Cloud also told the "China Times" reporter that the teams of the two sides have also completed the adaptation of the 4 billion parameter large model of Tongyi Qianwen and the Dimensity 9300, and will also adapt more large models such as 7 billion based on the Dimensity in the future to "proof" and support the development of more AI agents and applications. "It's not that the bigger the parameter, the better, but the most important thing is adaptation. A relevant person from Alibaba Cloud told reporters.

According to the reporter's understanding, Tongyi Qianwen is a basic large model developed by Alibaba Cloud, which has so far launched version 2.0 of 100 billion parameters and open source 72 billion, 14 billion, 7 billion, 4 billion, 1.8 billion, 500 million parameters and other sizes, as well as multi-modal large models such as visual understanding model Qwen-VL and audio large model Qwen-Audio.

When communicating with reporters, some industry insiders believe that the use of terminal computing power for AI inference can greatly reduce the cost of inference, ensure data security and improve the AI response speed, so that large models can better provide users with personalized experience.

On March 24, Lin Dahua, a leading scientist at the Shanghai Artificial Intelligence Laboratory, also proposed at the 2024 Global Developer Pioneer Conference Large Model Frontier Forum that with the exponential growth of cloud-side computing, device-side computing will usher in a golden period of growth. At present, many models are getting smaller and smaller, and they can be used directly in mobile phones. To this end, the computing power of mobile phones also needs to continue to grow. In the future, cloud collaboration will become an important trend, and the growth of large model users will mainly rely on the device side.

However, it should be mentioned that it is not easy to deploy large models on mobile phones.

According to the reporter's understanding, in order to deploy and run the large model on the terminal, it is necessary to complete the in-depth adaptation of software and hardware from the underlying chip to the upper-level operating system and application development, and there are many challenges such as the technology is not connected, the operator is not supported, and the development needs to be improved.

Xu Dong, head of Alibaba Tongyi Lab business, said that Alibaba Cloud and MediaTek have cooperated in multiple dimensions such as model slimming, tool chain optimization, inference optimization, memory optimization, and operator optimization to achieve efficient heterogeneous acceleration based on AI processors, truly "load" large models and run them in mobile phone chips, and successfully prototype a new Model-on-Chip deployment model for end-side AI for the industry.

Stirring up the sluggish mobile phone market

As the semiconductor company with the world's largest smartphone chip shipments, the cooperation between MediaTek and Alibaba Cloud on the device-side large model has a self-evident impact on the mobile phone market. According to the latest data from Canalys, it shipped more than 117 million units in the fourth quarter of 2023, with Apple in second place with 78 million shipments and Qualcomm in third place with 69 million shipments.

In fact, not only MediaTek, in March this year, Qualcomm also announced the launch of the third-generation Snapdragon 8s mobile platform, which supports large language models at the level of 10 billion parameters, and can also support multimodal generative AI models, which was launched by Xiaomi Civi 4 Pro for the first time in the world.

Mobile phone terminal manufacturers deployed AI large models earlier. According to the reporter's understanding, the Xiaomi 14 and 14Pro released in October last year have deployed AI device-side large models. In February this year, OPPO released its own 1+N agent ecological strategy, and Meizu is also announcing a strategic adjustment, stopping the development of new projects of traditional "smartphones" and investing in a new generation of AI devices.

Behind the all-out AI in the mobile phone industry chain, data from TechInsights shows that the replacement rate of smartphones in the world may fall to the lowest point of 23.5% (51 months replacement cycle) in 2023.

The "AI mobile phone white paper" released by IDC predicts that the new generation of AI mobile phones will drive a new round of replacement from 2024, and the global shipments of new generation AI mobile phones will reach 170 million units in 2024, accounting for about 15% of the overall shipments of smart phones, specifically for the Chinese market, IDC also expects that with the rapid iteration of new chips and user use scenarios, the new generation of AI mobile phones will reach 150 million units in 2027, with a market share of more than 50%.

Prior to this, the mobile phone market had been in the doldrums for far too long. In the past 2023, IDC data showed that global smartphone shipments fell by 3.2% year-on-year to 1.17 billion units, the lowest in a decade, and China's smartphone shipments that year were about 271 million units, down 5% year-on-year, and also set the lowest shipment volume in the past 10 years.

The AI model is expected to be a shot in the arm to stimulate the growth of the mobile phone market. However, Guo Tianxiang, a senior analyst at IDC China, pointed out to the China Times that with the upgrade of chips, there will be more mobile phones that can support large models in terms of computing power, but for consumers, AI large models will not become a rigid demand in a short period of time, and it is difficult to become the main thrust to promote consumers to change their phones.

Editor-in-charge: Huang Xingli Editor-in-chief: Han Feng