In addition to the ability to roll images, what else can a mobile phone roll?
The OPPO Find X7 series, the first flagship smartphone to be released in 2024, gives the answer - device-side generative AI.
Find X7 is available in 4 colors: Smoke Cloud Purple, Sea and Sky, Desert Silver Moon, and Starry Sky Black.
On January 8, 2024, OPPO released the new flagship Find X7, with a new generation of concentric universal design, a new super light and shadow triple main camera to achieve Hasselblad full-range master image, OPPO's self-developed tidal architecture released by the super performance, and a new generation of ultra-bright and ultra-eye-friendly diamond screen, once again revolutionizing the industry's flagship standard.
The OPPO Find X7 Uluta is the first to feature dual periscopes and four main cameras to form a Hasselblad Master lens group, once again refreshing the ceiling of mobile images. Liu Zuohu, chief product officer of OPPO, said that the word ultra stands for - the extreme, there is no intermediate state.
Imaging capabilities are already leading, and OPPO is also the first to show consumers what a generative AI phone looks like, and what is the difference between the world's first mobile phone with a 7 billion parameter large language model on the device side?
The AI function of mobile phones is not new, AI beautification and AI assistant have been widely used, but there are two obvious differences with the functions realized by generative AI, more intelligent and more diverse gameplay.
How to upgrade the AI experience of large models on mobile phones?
In terms of AI** elimination, traditional AI retouching is not intelligent enough and the effect is not ideal.
For example, if you take a landscape photo in a crowded scenic spot, and want to eliminate the passers-by in the **, the traditional AI practice has to manually select the range carefully, and after waiting for a long time, you can get the ** of the passers-by who are simply blurred, and the perception is very poor.
If you use the large model capability of the mobile phone to roughly frame the range that needs to be eliminated, you can easily eliminate passers-by in the landscape photo, and the AI will also "generate" the road surface to get a landscape photo without crowds.
This is the AIGC elimination function of Find X7, which is different from traditional AI models that only support human recognitionThe large model capability of Find X7 not only supports the identification and segmentation of more than 120 types of subjects, but also realizes hair-level segmentation and up to 6 multi-subject separationsAt the same time, it can also realize the filling and natural generation of large-area images, which is the experience of image editing in the era of generative AI on mobile phones.
Just as popular as AI editing is AI voice assistants.
Most people still feel that the existing voice assistants are not smart enough to use, and they will be forgotten as soon as the freshness passes. However, with the AI model on the mobile phone, the experience has also been improved by leaps and bounds, and the AI assistant can better understand natural language, and the conversation is more like a chat between people.
With the blessing of the AI large model, OPPO's new Xiaobu can not only chat, but also answer various questions, with a total of more than 100 capabilities, including text generation**, text generation, AI article summarization and other functions, which can help users in different dimensions such as office efficiency, life services, learning and education.
For example, you can ask Xiaobu why there are many small holes in the biscuitsAnd let Xiaobu answer in a way that a four-year-old child can understand.
In addition to upgrading the existing AI experience on mobile phones, generative AI models also bring new AI capabilities to mobile phones.
For example, at the press conference of the Find X7 series, Liu Zuohu showed a voice summary function of the AI large model after the exchange with Li Xiang**, CEO of Li Auto, which clearly presented the key points of the call and generated to-do, which greatly improved the efficiency.
It's got to beIt is beneficial to the 7 billion parameter model on the device side of Find X7, compared with other 1 billion parameter models, with a higher "IQ" comprehension abilityConversations can be understood more accurately and summarized with focused and detailed details can be generated.
At the same time, with the help of the intelligent scheduling of the device-cloud collaboration of the OPPO Andean large model, the whole process of processing the voice summary of the OPPO large model is encrypted, and the generated summary content is completely stored locally on the terminal, which can well protect user privacy.
Privacy protection and faster speed are the advantages of the device-side large model.
Compared with other models on the same platform, the 7 billion parameter large language model on the device side of Find X7 can bring 20 times faster response in 200-word cap generation, and 25x faster.
In addition, the 7 billion model on the device side of the Find X7 can also achieve a maximum of 14,000 abstract word limits, reaching 35 times, the dialogue experience of the large model is closer to the speed and capacity of human dialogue.
How to make the world's first mobile phone with a 7 billion parameter large language model on the device side?
Find X7 is the world's first mobile phone to apply a 7 billion parameter large language model on the device side, and through the deployment of device-cloud synergy, it has achieved a leading AI experience of more than one generation.
Why are we the first to really use the 7 billion model on the device side?How to adapt and deploy models, algorithms, GPUs, and NPUs is not a simple matter. Liu Zuohu further explained, "This process requires close communication with the platform factoryBased on our understanding of chips and insight into users, it is our core competitiveness to customize chips according to large models and algorithms, and to provide requirements to chip manufacturers in the language of chips.
Liu Zuohu revealed,".We've retained some of the core architectsThe purpose is to cooperate with platform manufacturers, such as Qualcomm and MediaTek's deep customization chips. We don't make chips ourselves, and the chip platform factory is very willing to work with us to achieve a strong alliance. If you don't understand chips and haven't made chips, you can only stay on the 'water floating' and can't enter the bottom layer, such as our self-developed tidal architecture, which is more valuable than redoing it.
Tidal architecture is a collection of OPPO's self-developed chip software and hardware fusion technology, and it is a chip optimization technology created by a group of people who understand chips.
The tidal structure deeply touches the SoC's L3 cache and SLC (system-level cache), and the SLC of the traditional SoC often adopts a fixed allocation scheme for computing cores such as CPUs and GPUs, which will cause certain resource waste and congestion.
The tidal architecture can flexibly adjust the cache ratio according to different application scenarios.
For example, when running a large game requires image rendering, GPU allocation is more, reducing the frequency of DDR and UFS access, improving the application running speed, and reducing the overall energy efficiency of the computing link. It has been calculated that the scheduling optimization of SLC can bring an average of 8% energy efficiency savings.
Tidal architecture is indispensable for the deployment of device-side large models, but device-side large models face more extreme computing and storage challenges.
Zhang Jun, general manager of OPPO Software Innovation Center, explained that the normal model size of the 7 billion large model is 28GB, in order to truly deploy on the device side, OPPO uses high-precision 4 bits to compress and lightweight the model, and the model is compressed from 28GB to the smallest 3About 9GB.
Even so, it is still a great test for the performance, memory and power consumption of the mobile phone.
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 situation of the machine, the power consumption of our device-side large model is controlled within the acceptable range of the user. Zhang Jun said.
The OPPO Find X7 series deploys a large language model with 7 billion parameters on the device side, and the vision model will also be deployed on the device side, which will be launched with our OTA later. Zhang Jun emphasized, "WeTruly deploy the 7 billion parameter model on the end-side of Qualcomm and MediaTek chip platforms at the same timeOPPO is the first. ”
As for why the 7 billion parameter model was chosen, Zhang Jun also gave an explanation,"Why choose a model with 7 billion parameters instead of 1 billion parameters?We choose the most appropriate parameter size for each scenario, which represents how smart the model is in this scenario. ”
In addition to technical capabilities, the key to success in the mobile AI field is strategy.
I think the real focus for the next few years is product-driven. Based on what products to be made, determine the training of large models, and the underlying computing power, and the overall route is still product-driven technology. Liu Zuohu believes that "the most important way to make products through user insights is the most important." Whoever understands the user best and who really understands the product will be able to maintain differentiation in the competition and always have the core competitiveness.
Leifeng.com learned,OPPO set up an AI center last FridayAll AI-related employees of the company are gathered, and the goal is to concentrate all resources to do AI well.
OPPO has kicked off the era of generative AI for mobile phones, and what unexpected applications can the future device-side large model bring to mobile phones?