The last track that opened a years-long arms race for all mobile phone manufacturers was imaging, and the next such opportunity is AI.
Text: You Yong
Edited by Zhou Luping
On the morning of February 8th, OPPO officially released a poster to announce that something big would happen tomorrow......
Three days ago, OPPO released a preview of the backstage of the CCTV Spring Festival Gala rehearsal, and the host of the Spring Festival Gala, Nigmaiti, Ren Luyu, and Ma Fanshu specially experienced the AI functions of the OPPO mobile phone, including AI elimination, AI call summary and Xiaobu New Year's greetings.
OPPO will show the capabilities of AI in front of the people of the whole country through the stage of the Spring Festival Gala, and it is also announcing that the mobile phone industry will enter a new era of AI.
01
The curtain of an era is opening
The far-reaching impact of ChatGPT in the field of AI continues to ferment. If we say that the early stage is more of a carnival of cloud vendors, they have launched general-purpose large language models, and have begun to look for landing applications in thousands of industries.
In the past six months, with the maturity of the underlying chip platform and the advancement of model quantization and compression, hardware terminal manufacturers such as mobile phones and PCs have accelerated the pace to promote the application of large models on the device side.
At the end of last year, Qualcomm and MediaTek's latest flagship platforms both supported the deployment of device-side large models. For example, Snapdragon 8 Gen3, as the first mobile platform built specifically for generative AI, supports models running 10 billion parameters on the device side and generates up to 20 tokens per second for 7 billion parameter large language models. MediaTek's Dimensity 9300 is also equipped with the AI performance engine of the seventh-generation APU 790, which can support terminals to run AI large language models with 1 billion, 7 billion, 13 billion, and up to 33 billion parameters.
With the blessing of the underlying chip platform, OPPO, Xiaomi, vivo, Honor and Samsung have all launched or will soon launch 7 billion parameters of the end-side model. As an intelligent body, mobile phones play an increasingly critical role in the application of device-side large models. Large models are also considered to be the key to bringing a revolutionary user experience in the future, which in turn will lead to device replacement. Based on this understanding, it has become the strategy of mobile phone manufacturers to devote themselves to the research and development of AI.
It is reported that Apple, which has always been cautious, is also expected to introduce a series of AI experiences based on large models for products such as Siri at the Worldwide Developers Conference in June this year.
The PC industry is also evolving at full speed toward AI PCs. Intel CEO Gelsinger bluntly said, "AI represents the arrival of a new era. AI will fundamentally change, reshape, and reframe the PC experience through the close collaboration between the cloud and the PC, unleashing people's productivity and creativity, and we are moving towards a new era of AI PCs. ”
Intel has specially released the Core Ultra processor, which for the first time combines CPU, GPU, and NPU in one to meet the needs of AI PCs. AMD and Qualcomm have also successively launched CPUs with integrated AI modules. Based on Intel's AI processors, Lenovo launched the first notebook with AI capabilities, announcing a full transformation to AI PCs.
Huo Jinjie, President of IDC China, pointed out that AI PC is not only a new product, it represents that the entire industry is about to enter a new era - an era of great changes in the industry, an era of historical breakthroughs. And Counterpoint's view is that it expects AI laptops to share 50% of sales by 2025, with at least one NPU or AI accelerator to complement the capabilities of its main CPU and GPU.
Both PCs and mobile phones are industries that play an important role in daily life and work, but also face growth difficulties, and now they all invariably put AI in an extremely important strategic position.
Behind the arrival of the new era, the current PC and mobile phone industries are facing a lot of growth difficulties. According to the latest statistics from IDC, in 2023, China's smartphone market will ship about 271 million units, a year-on-year decrease of 5.0%, the lowest in nearly 10 years. There are many factors behind this, including the lack of innovation in smartphones themselves, and the increase in user replacement cycle from the previous two years to the current three or four years.
Today, the tide of the mobile phone industry is turning, and AI will be the best catalyst to stimulate industry innovation and sales growth.
2024 is considered the first year of AI mobile phones. An industry insider judged that the last track that allowed all mobile phone manufacturers to start a years-long arms race was imaging, such as more lenses, larger sensors, etc., and in the past few years, mobile phone imaging capabilities have been substantially improved. The next such opportunity is AI, but compared with imaging, AI is a longer journey, and the requirements for technology accumulation and R&D investment will be higher.
02
Seize the opportunity of large-scale mobile phones
"If you don't lay out a large model of mobile phone manufacturers, there will be no play!" Liu Zuohu, senior vice president of OPPO, said very bluntly and forcefully, and this view also represents the current general sentiment in the mobile phone industry.
According to Counterpoint's forecast, 2024 will be a pivotal year for GenAI (generative AI) smartphones, with shipments exceeding 100 million units, and by 2027, GenAI smartphone market share and shipments will reach 40% and 522 million units, respectively.
Although there is no unified standard for the definition of AI mobile phones in the industry, it also shows from the side that AI will be the standard configuration of future mobile phones, and this standard configuration is not just an independent application and scenario, but reconstructs the experience of mobile phones from the bottom of the system, so that AI capabilities are ubiquitous on the system side, "the future mobile phone is an AI Agent".
At the beginning of January this year, OPPO's annual flagship Find X7 series was released in Shenzhen. There is a big difference between this press conference and the previous ones, which started with the introduction of the image or the design of the mobile phone. But this time, OPPO spent nearly half an hour demonstrating the application landing brought by the large model on the end side. Other vendors have made similar adjustments at their press conferences, and each has raised the priority of the device-side large model to the most important position.
The OPPO Find X7 is the industry's first model to land a 7B device-side model. At the press conference, Liu Zuohu demonstrated several scenarios implemented by the device-side large model, such as call summary, which will generate a concise but focused summary based on the call record, and then generate to-do lists, and even help users record key points when they don't have time to answer the phone, so that no important information will be missed in work and life scenarios. Moreover, compared to the 1B model, the 7B model performs significantly better in abstracts.
In addition to the language model, OPPO's AndesGPT also has a comprehensive and powerful generative vision model, which can accurately understand the semantics of images, and on this basis, it realizes the recognition and segmentation support of more than 120 types of subjects.
In addition to OPPO, Samsung has also begun to explore the application of large models on the end side. Compared with OPPO's call summary capability, the Galaxy S24 shows the scene of call translation, and the two parties can communicate in real time across languages, which is equivalent to having a simultaneous interpretation around them, which is eye-catching.
Overall, Samsung and OPPO are relatively similar in the application direction of the device-side large model, and both show AI applications in cutout and call scenarios. However, the implementation paths and emphases of the two are different, for example, Samsung's ability to use large models on the device side mainly comes from Google's Gemini Nano model, while OPPO is its self-trained large model.
A person in the mobile phone circle told the front line of digital intelligence that most of the functions of the manufacturers on the market that claim to have realized the 7B device-side large model are actually implemented in the cloud, because the speed of the cloud is fast enough, and the end-side is limited by factors such as power consumption and memory. According to him, only OPPO of domestic mobile phone manufacturers has really achieved the effect of a large model on the end side, one is call summary, the other is intelligent elimination, and other manufacturers are basically processed in the cloud.
At present, the biggest problem in the deployment of large models on the device side lies in the compression and quantization of the model and the improvement of storage performance on the device side.
According to another person familiar with the mobile phone manufacturer, the biggest bottleneck of the device-side large model is not in the GPU or NPU, but in the memory bandwidth. At present, the actual bandwidth of the most powerful memory LPDDR5X is more than 50 GB per second, and the 7B model accounts for about 4G memory after quantization and sparseness, and its performance needs to reach more than 10 tokens and more than 40 GB per second, which is exactly the bandwidth limit of LPDDR5X.
In order to realize the landing of the large model on the end side, OPPO has reduced the memory occupation of the 7B large model to less than 4GB, the performance has reached more than 10Token/S, and the power consumption has now reached the 800 mAh level, "It is acceptable in this kind of scene and will not make the phone hot." It is understood that the OPPO Reno series of products will also land on the end side of the large model.
However, the device-side large model deals more with some simple tasks, and most scenarios are still inferred in the cloud. Device-cloud collaboration is currently the mainstream of large models in mobile applications.
Therefore, in addition to the 7 billion parameter device-side large model, OPPO has also trained the 70 billion parameter AndesGPT-Turbo and the 180 billion parameter AndesGPT-Titan, which can meet the deep inference of highly complex tasks, and both models are deployed in the cloud.
According to OPPO, there are two directions for the technical evolution of large-scale model mobile phones in the future: one is to continue to tackle the end-to-side deployment of large-scale models to achieve better performance and stronger processing capabilities, and the other is to use a more general and advanced architecture to solve the problem of device-cloud collaboration.
03
Winter-resilient innovation for good
Any layout that seems forward-looking now doesn't happen in a vacuum. In fact, OPPO's investment in the research and development of large models is not short.
As early as 2020, OPPO has launched the exploration and practice of pre-trained language models, and self-developed large models with 100 million, 300 million and 1 billion parameters.
However, both BERT and GPT are based on the Transformer architecture, which also lays a solid foundation for OPPO for the subsequent AndesGPT training and has accumulated rich experience. Today, AndesGPT excels in general Q&A and conversation, and has topped the C-Enal global Chinese list.
OPPO is increasing investment in technology to ensure that it does not fall behind in the competition in this red sea market, which has also become the most effective magic weapon for OPPO to resist the cold winter of the industry.
If it is said that OPPO used to spend a lot on marketing, but an intuitive change in recent years is that OPPO is more and more willing to invest in technology and product research and development. The outside world has also seen more and more OPPO's innovation leadership, including the industry's first device-side large model, the industry's first dual periscope flagship, and breakthroughs in communication and chip underlying technology. Under the concept of innovation for good, OPPO is transforming from a single mobile phone company to an ecological technology company.
Not long ago, OPPO and Nokia officially announced the signing of a 5G patent cross-licensing agreement, after which the two parties will end all pending litigation in all jurisdictions. The nearly three-year-long patent dispute ended dramatically in a settlement.
The significance of this matter is not limited to OPPO itself, but also gives a good example to domestic enterprises in the face of unreasonable patent licensing.
Of course, strength is the strength to confront unreasonable. OPPO's huge investment in R&D and the accumulation of a large number of patents over the years have escorted its global business and security operations. At present, about 60% of OPPO's employees are R&D personnel. As of December 31, 2023, OPPO has filed more than 98,000 patent applications worldwide. Among them, the number of invention patent applications exceeded 89,000, accounting for more than 90%, ranking 6th in the world, second only to Huawei in China. In these places where the outside world is not easy to pay attention to, OPPO is accumulating energy.
It is the investment in R&D and the accumulation of patents over the years that have given OPPO the confidence to choose to face up to international manufacturers. It is reported that with the reconciliation between OPPO and Nokia, it is also paving the way for OPPO to return to some European markets in the future.
And these technologies and patents are ultimately used in the product experience. Taking the newly released Find X7 series as an example, in addition to the above-mentioned implementation of the device-side large model, there are also optimizations in communication technology, chip underlying architecture, and security.
The "Tidal Architecture" released this time solves the problem that the storage and read and write speed of the SoC is much lower than the computing speed. The Tidal architecture optimizes the chip performance down to the L3 cache level, and dynamically regulates system resources by identifying the needs of application scenarios. For example, if there are many computing tasks, more cache will be allocated to the CPU, and if there are many rendering tasks, resources will be tilted to the GPU. This architecture enables SoCs to save an average of 8% in energy efficiency.
In addition, OPPO's complete set of self-developed communication technology solutions, Super Signal Engineering, includes the overall breakthrough of antenna technology, smart communication, and near-field communication, which can also provide "the best communication and network connection experience in the user's scenario".
Today, AI and imaging have become the two pillars of OPPO's technology. Liu Zuohu revealed that OPPO recently gathered all the company's AI-related partners together to set up an "AI center", and the group's technical leader independently managed and coordinated. In the past, these positions were scattered in various departments and were not systematic, "We decided to pool all our resources and set up an AI center to make AI the core competitiveness of OPPO." Liu Zuohu said.
From the current point of view, the application of large models on the device side has just begun. In the future, whether there is a large model and how to use it in depth will determine whether the product is competitive.
Of course, as Bill Gates famously quotes oft-time – we always overestimate what will happen in the next two years and underestimate what will happen in the next decade. This sentence is also applicable to AI, whether it is an AI mobile phone or an AI PC, in fact, it is still in the initial or preparatory stage, and it needs to go through a relatively long period of iteration and evolution before the real revolutionary experience can be realized. But there is no doubt that 2024 has become a recognized important moment for AI mobile phones.
There is a saying in "Flowers" that is very apt: the goal is never far away, step by step, day by day. Just give it your all, and leave the rest to time.