In the era of generative AI, mobile phones are undergoing a full stack revolution?

Mondo Technology Updated on 2024-03-05

Released by the Heart of the Machine.

Editor: Zenan

The third major change in the mobile phone industry began.
In recent times, AI and large model technology have advanced by leaps and bounds. Just after the Spring Festival, a new round of breakthroughs was ushered in in the frontier direction.

OpenAI's SORA suddenly slowed up the progress bar generated by AI ** by half a year.

In the application field of large models, the speed of technology application is also accelerating. At present, the new generation of flagship mobile phones from major manufacturers has all appeared, and most of them are equipped with large models, which can achieve many unprecedented functions.

*From the Qualcomm Snapdragon 8Gen3 trailer:

Why did everyone choose to enter AI in 2024?

The answer seems clear. With the gradual maturity of large models and the enhancement of chip-side computing power, mobile phone manufacturers have a clear judgment: 2024 will be the first year of AI mobile phones.

However, in the face of the same goal, each company has different ways of playing: many mobile phones have begun to introduce cloud large-scale model applications, and some have achieved end-to-end running of small-size models. Among them, the OPPO Find X7 series, which has landed multiple AI functions, is gaining more and more recognition.

It has achieved many firsts: Find X7 is the world's first mobile phone to apply a large language model with 7 billion parameters on the device side.

It's also the first phone to support AI call summaries, allowing users to have AI generate summaries with one click after a call ends, and automatically generate to-dos and reminders

With the blessing of the AI model on the mobile phone, the smart assistant is no longer a decoration, and OPPO's Xiaobu assistant has achieved a leapfrog experience improvement.

Now Xiaobu is able to understand natural language better and can answer all kinds of tricky questions. It has more than 100 capabilities, including text generation**, interpretation, AI article summarization, and more, which can help users in different dimensions such as office efficiency, life services, learning and education, and more.

For example, you can ask Xiaobu why there are many small holes in the biscuit, and ask it to answer in a way that a four-year-old can understand

OPPO has also brought a brand new AIGC elimination feature. In the past, the redrawing task required the complex operation of professional computer software, but now you only need a simple circle of AI on your mobile phone to help you complete the subject elimination and real-life redrawing. In this process, the large model can also carry out a certain degree of creation, the brain makes up the background, and gets a landscape photo without crowds

Find X7's large model capability not only supports the identification and segmentation of more than 120 types of subjects, but also enables hair-level segmentation and up to 6 multi-subject separations. This is the new experience of taking photos in the era of generative AI on mobile phones.

As the most common consumer electronic product, mobile phones have always been the priority direction of various AI technologies, and functions such as AI beautification and AI assistant have long been the standard configuration of smartphones. And with the rise of large models, mobile phones as human epitaxial organs to understand the advantages of users, coupled with the unprecedented breakthrough of generative AI, have brought a smarter, more personalized experience and more diverse gameplay.

AI mobile phones may be the third major stage of change in the mobile phone industry after feature phones and smartphones. In this change, OPPO showed us in advance the new form of mobile phones after the breakthrough of large-scale model technology.

7 billion parameter large model on the device side

Why did OPPO do it?

Many of the new capabilities that the OPPO Find X7 can achieve are due to the 7 billion parameter model running on the device.

For the first time on this generation of flagship phones, OPPO's self-trained AndesGPT 7 billion parameter large model is equipped. Through the deployment of device-cloud synergy, it has achieved more than a generation-leading AI experience. Compared to the 1 billion parameter model, the model exhibits a higher IQ comprehension, which can more accurately understand the content of the conversation and produce focused, detailed summaries.

Compared to other models on the same platform, the AndesGPT 7 billion parameter version can bring 20 times faster response when generating 2000 words, with a maximum of 140,000 words for a content summary, which is 35 times, the conversation experience is also closer to the speed and amount of information of human conversations.

We know that in the current large-scale model arms race, various technology companies are rushing to buy GPUs, after all, running large models consumes computing power. It's not easy to run a large model on a phone with limited computing power and memory, but how does OPPO do it?

This is because on the new generation of mobile phones, OPPO has achieved a comprehensive optimization for large models, from software, hardware to cloud platform.

First of all, OPPO worked closely with platform manufacturers, based on the understanding of the chip and the insight and understanding of user needs, deeply customized the SoC chip according to the large model and algorithm, improved the operation scheduling of the chip under high load conditions, and then optimized the operation efficiency of the large model.

Insight and understanding of user needs is OPPO's core competitiveness, and Call Intelligence Summary is a good example: Call recording is a feature that has been around in the industry for a long time, but combining traditional call recording with a large device model brings a new AI experience that subverts traditional applications.

In addition to the extreme optimization of computing resources, another bottleneck is storage.

Under normal circumstances, a large model with 7 billion parameters would take up 28GB of memory. In order to truly achieve device-side deployment, OPPO has greatly compressed the model in the way of int4 quantization, so that the model that originally occupied 28GB of memory now only needs 39GB, which reduces resource requirements and barely affects the output of the AI model.

The performance consumption of large models on the device side is relatively large, so it is necessary to optimize the operator of parallel computing and the optimization of memory management to reduce the loss and system resource occupation. In terms of battery life, according to the use of the machine, the power consumption of our device-side large model is controlled within the acceptable range of users, said Zhang Jun, product director of OPPO AI Center.

OPPO is also the first mobile phone manufacturer to deploy a 7 billion parameter model on Qualcomm and MTK platforms at the same time, which not only treats users equally, but also verifies its ability to optimize the deployment of AI. At present, the mobile deployment of the device-side vision model is also on their schedule.

For generative AI capabilities, sometimes complex tasks that cannot be handled by device-side models require data transfer to the cloud and leverage the power of server-side AI accelerators. A lot of information containing personal information and preferences needs to be pre-processed on the mobile phone side to ensure privacy.

The device-cloud collaboration of large models is the consensus of the current industry.

In this regard, when OPPO released the AndesGPT large model, it proposed a device-cloud joint deployment realized through the first-class large model deployment strategy, which met the efficient adaptation of multiple scenarios.

Under the architecture of device-cloud collaboration, the computing power supply of OPPO AI mobile phones is no longer limited to local, and user information can also be guaranteed to be not leaked - user data is only calculated on the device side, and the more powerful computing power in the cloud is oriented to complex task processing, which not only improves the overall performance and efficiency of large model calculation, but also ensures security.

In addition to AI technology capabilities, OPPO also has a big strategy.

Stepping up your AI strategy

There is no limit to the investment in artificial intelligence

In fact, OPPO has been deeply engaged in the AI field for many years: as early as 2020, OPPO has begun to explore the training, application and landing of large language models, and the first self-developed large model OBERT once jumped to the top of the Chinese large-scale knowledge graph Q&A kgclue ranking; In 2023, OPPO's self-trained Andean model (AndesGPT) is second only to GPT 4 in the Super Clue Knowledge and Encyclopedia Ability Rankings, ahead of all competitors.

In addition, the 7 billion large models applied on the device on Find X7 also ranked first on the Android list.

Chen Mingyong, founder and CEO of OPPO, judged that 2024 will be the first year of AI mobile phones, and the impact of AI on the mobile phone industry in five years will not be negligible.

On February 20, OPPO held an AI strategy conference to share their detailed plans for AI strategy.

OPPO defines a true AI phone with four characteristics:

Efficient use of computing resources to meet the computing needs of generative AI;

Be able to perceive the real world keenly and understand the complex information of users and the environment;

Strong self-learning ability;

Possess more abundant creative ability to provide users with continuous inspiration and knowledge support.

In order to meet these capability requirements, the mobile phone industry needs to carry out comprehensive technological innovation and ecological reconstruction: in terms of hardware architecture, high-performance AI computing power base, management optimization of model library, and intelligent bionic perception capabilities will become the new standard for AI mobile phones. The AI phone's OS system will be able to efficiently handle complex tasks and create proactively by embedding agents.

On the other hand, future AI phones will support sharper and more accurate understanding of natural semantics, with more powerful self-learning capabilities, and can provide more intuitive multimodal interactions. From this point of view, the traditional application ecosystem will shift to the intelligent twins ecosystem in the era of AI mobile phones, and all kinds of service applications will be seamlessly connected with AI capabilities to achieve truly intelligent services.

For this, OPPO is well prepared. In terms of cloud computing power, OPPO has an OPPO AI Marina Bay data center that can support hundreds of billions of AI model training, ultra-low latency for backbone network links of 2 milliseconds, and 100% pure green energy.

In the future, OPPO will continue to invest in computing power and deploy models at three levels: AndesGPT Titan, Turbo, and Tiny to correspond to different application scenarios.

In terms of agent capabilities, OPPO officially released the 1+N agent ecosystem strategy. One of them stands for the OPPO AI superagent, which is based on knowledge graphs, document data, and search engines, which can accurately understand user intent, give accurate results, and fully invoke a variety of other tools. n represents a new agent ecosystem based on the OPPO AI Pro agent development platform.

Compared with large-scale intelligent assistants, intelligent twins are faster and more proactive AI assistants, which can complete various goals in the most effective way according to your settings, and meet the needs of various scenarios such as emotional companionship, knowledge exploration, and entertainment and chatting. At the same time, the method of building an agent is very simple, no programming foundation is required, people only need to have a natural language conversation with the large model, and provide the necessary instructions and knowledge.

The agent can also connect with other services outside the large model to access more information and mobile phone functions, and meet the various needs of users with generalized capabilities.

What's more, OPPO's 1+N also means uniting more partners and developers. With the development of the intelligent twin, people can jointly create a service system for the AI mobile phone ecosystem.

It's like an app market on a smartphone. But this time, the services connected by AI are more capable and connected to you.

Oppo continues to ramp up AI. At the strategy conference, OPPO officially announced the establishment of an AI center, aiming to integrate R&D resources and conduct capacity building and R&D for AI. Liu Zuohu said that the establishment of the AI center will bring together the strength of the entire company, and AI has been regarded as the most important strategy in the next era of mobile phones, and there is no upper limit on investment.

A mobile phone blessed by a large model

How else will it evolve?

And that's just the beginning. After all, the big model is a world-changing technology, and applications in all fields have to be redone with AI.

Recently, there has been a lot of big moves in AI: Apple has halted its multi-year car-building project and shifted its focus to generative AI; When Google released the native multimodal large model Gemini, it announced that the large model will be integrated into the Android system in the future; The next-generation AI Hub launched by Qualcomm at MWC has supported the acceleration of more than 75 mainstream AI models on the device side. Now, from mobile phone manufacturers to technology companies to chip companies, heroes see the same thing.

We can foresee that with the advancement of technology and the construction of industry ecology, we will also see more intelligent photography, faster human-computer interaction, more personalized content generation and more efficient task processing in the future.

With generative AI, the complex work of the past will be much simpler. Without having to open professional software, people can let AI automate complex tasks with just verbal instructions, greatly improving productivity.

Generative AI deployed on the device side can also make smartphones more aware of users' habits and location. With contextual information, digital assistants will be more personalized, with more satisfying answers and more proactive service.

Furthermore, as AI generation capabilities move into the multimodal realm, next-generation AI rendering tools will be able to generate 3D objects and scenes using various types of prompts such as text, voice, images, or **, ultimately creating new immersive content experiences.

In a word, AI phones are about to revolutionize us.

With the continuous promotion of AI technology by mobile phone manufacturers such as OPPO, the distance between us and imagination has become closer.

Related Pages