HomeAI News
The new battlefield of “Huami OV”: large model of mobile phone
81

The new battlefield of “Huami OV”: large model of mobile phone

Hayo News
Hayo News
October 26th, 2023
View OriginalTranslated by Google
The trend of big models reshaping everything has spread to the mobile phone industry.

After witnessing the reshaping of IT infrastructure, marketing, office, autonomous driving and other industries by large models, mobile phone manufacturers, which are deeply trapped in the "trough" of the double cycle of sales and technology in the mobile phone industry, urgently and high-profile poured into large models + mobile phones. track.

Leading the way are Huawei and Xiaomi. In August this year, Huawei and Xiaomi successively announced the integration of their own mobile assistants into large models. For a time, the mobile assistant also had in-depth conversation, natural language understanding and programming capabilities similar to ChatGPT.

However, such a large model deployment method that relies on the cloud requires users to wait for a long time, consumes a lot of bandwidth and computing resources, and the dialogue is often not smooth.

Therefore, the solution of deploying large models on the mobile phone to form a large model of device-cloud collaboration has become the unanimous choice of all mobile phone manufacturers.

In order to put large models into mobile phones, OPPO, vivo, and Xiaomi have successively launched "lightweight versions of large models" with a parameter level of one billion for mobile phones. Chip manufacturers MediaTek and Qualcomm have also launched mobile phone SoC (system-on-chip) that are more suitable for AI computing based on the needs of large models.

After the upcoming OPPO and vivo developer conferences in November, "Huami OV" people will gather together again on large mobile phone models.

Nowadays, although mobile phone manufacturers are still in their infancy, they have reached consensus on how to make good use of the capabilities of large models, such as: how to compress models, how to collaborate between cloud and terminal, and how to connect large models and systems and applications at the bottom layer. capabilities, etc.

After years of hardware evolution, mobile phones have once again reached a new node of software-defined mobile phones, which further tests the research and development capabilities of mobile phone manufacturers. Because of this, mobile phone manufacturers have recently released and upgraded their own self-developed operating systems in order to outperform the software experience.

It is foreseeable that a wave of new selling points of large models➕mobile phones is on the way. As the computing power requirements of large models increase, can it trigger a new wave of replacements and pull mobile phone manufacturers out of the cold winter?

Focus on the new battlefield of the large model "Huami OV"

Chinese mobile phone manufacturers like to "fight as a team", and their entry into large models is no exception.

After "Huami OV" is in place, how will the large model be implemented on mobile phones? We are about to enter the moment of “practice tests the truth”.

The first is the vivo Developer Conference that is just around the corner. As early as September, Vivo Executive Vice President and Chief Operating Officer Hu Baishan announced that he would release a large self-developed model and would see you together with the new phone.

However, it may be that "the news is too big to hold it in". Only a month has passed, and around the vice president of vivo and the director of the Global AI Research Institute, people have begun to continuously "spoiler" the information and capabilities of the upcoming large model on Weibo. .

Among them, the soon-to-be-released vivoLM (vivo large model) is divided into 5 versions, which are divided into three levels according to the number of parameters: billion (1B/7B), tens of billions (66B) and hundreds of billions (130B/175B). Among them, version 7B is the version that vivoLM will open to the outside world.

In terms of specific capabilities, large models are good text generation assistants and are also good entrances to knowledge acquisition. People think, "Text generation assistant, in addition to ordinary writing, summarizing, and expanding, it can even help you write SWOT analysis. At the same time, it is also a good entrance to acquire knowledge. As long as you ask the right questions, it can be used by everyone." Comprehensible sentences explain knowledge in many professional fields, which is much more efficient than searching for gold in search engines. It is truly 'once you use it, it's hard to put it back'."

According to this description, the current combination of large models with mobile phones may be closer to the image of a smart assistant.

On October 11, OPPO announced that a new version of its AI assistant based on the AndesGPT large model, Xin Xiaobu 1.0, has launched its first round of public testing.

OPPO’s vision for the application of large mobile phone models is also the image of an assistant. "Today AIGC is really like a person talking to you. This is the most unimaginable experience. Therefore, the mobile phone of the future must be your super assistant." Regarding the disruptive nature of AIGC and large model technology, OPPO Senior Vice President President and Chief Product Officer Liu Zuohu said in a recent interview.

Compared with "OV"'s "rush" in November, Huawei and Xiaomi have already started applications on their respective mobile assistants.

On August 4, at the Huawei Developer Conference, Huawei Managing Director, Terminal BG CEO, and Smart Car Solutions BU CEO Yu Chengdong announced the launch of a new Xiaoyi with large model support. The updated functions are mainly reflected in smart interaction, which can complete natural Functions such as device control for language understanding, copywriting content assistance, and image secondary creation.

On Xiaomi's side, on August 14, Xiaomi launched an invitation test for the XiaoAi large model, covering a large number of mobile phone models and some smart speakers. The upgraded version of Xiaoai has the capabilities of contextual understanding, higher-quality Q&A, and generative content output. It is basically equivalent to embedding the capabilities of large-model apps such as Wenxinyiyan and iFlytek Spark directly into mobile phones. But different from App, this kind of embedding will be lower level and the interactions will be more diverse.

How to fit large models into mobile phones?

Putting a large model into a mobile phone is like putting an elephant into a refrigerator. There are three steps.

First of all, although the mobile phone industry has different views on the specific architecture and path of large-scale model applications, a consensus has been reached on the collaborative design idea of ​​"cloud + terminal".

In data centers, large models often have tens or hundreds of billions of parameters, which correspond to the hardware requirements of tens of thousands of GPUs during training. It is impossible to fit a complete large model into the "consumer-grade" hardware of a mobile phone.

Under the pure cloud solution, the user's waiting time is too long due to the superposition of network communication time, cloud computing transmission time, and judgment feedback time, and the experience is very poor. It takes up to two seconds to recommend words using the input method combined with a large model. It cannot be used in real-time chat scenarios, let alone in locations with weak signals such as basements, elevators, and old office buildings.

In addition, if you want to provide more intelligent assistance, you will inevitably come into contact with more user information. If all this information is uploaded to the cloud, privacy and other information security cannot be guaranteed. Even if information security can support it, all this information will be uploaded to the cloud. The huge amount of data generated by the current huge stock of smart hardware will be a huge problem for manufacturers who are already tight on GPU hardware. This will bring further pressure on hardware such as network bandwidth and storage.

As Qualcomm Senior Vice President Alex Katouzian said, “As connected devices and data traffic accelerate, and superimposed data center costs rise, (we) cannot send everything to the cloud.”

As a result, if you want mobile terminals (mobile phones) to have large model capabilities, it is necessary to arrange end-side large models on the end side to "respond".

The layering of cloud and terminal is the first step in terminal-cloud collaboration.

"Model training requires huge computing power and must be carried out in the cloud. The application on the mobile phone is actually inference. During inference, the model can only activate part of the modules and part of the neurons to calculate, " Xiaomi Technology Committee AI Experiment Luan Jian, head of the large model team of the laboratory, said.

In addition, the specific application scenarios of large models on the cloud and on the device are also different.

For example, you can perform some simple tasks on your mobile phone, such as writing, making suggestions, drawings, etc. Others involve complex knowledge systems and real-time information that are not suitable for data collection and learning, such as ticket booking, hotel booking and other operations, which can mobilize the capabilities of the cloud.

The next step is to transform the mobile phone SoC and add GPU, NPU (neural network processor), APU (acceleration processor) and other hardware suitable for large models based on the original hardware platform.

On the new flagship chips, Qualcomm and MediaTek have made corresponding upgrades for large models. For example, on the Snapdragon 8Gen3, which was released on October 25, the NPU performance increased by 98%, and two low-power NPU units were provided. Surprisingly, Snapdragon 8Gen3 supports a 10 billion parameter model. This means that Qualcomm’s current new hardware is more than enough to accommodate the large models of mobile phone manufacturers.

On the other hand, MediaTek, which has in-depth cooperation with vivo and OPPO, has also integrated a new AI processor on its upcoming Dimensity 9300 chip.

However, hardware upgrades are not enough. If you want to "pack" a large model into a palm-sized mobile phone, you also need to "slim down" the large model.

At the parameter level, mobile phone manufacturers generally choose billion-level large models for implementation. For example, Xiaomi’s 1.3 billion, vivo’s 1 billion and 7 billion, OPPO’s 7 billion, etc. The reduction in the number of parameters effectively reduces the storage space occupied by large models on the side, and avoids the embarrassment of hundreds of billions of large models occupying hundreds of gigabytes of storage space just for parameters alone.

After the parameters are streamlined, manufacturers need to further adjust the model so that small parameters can also produce results with large parameters.

Take Huawei's new Xiaoyi as an example. Xiaoyi is a dialogue model built and fine-tuned for end consumer scenarios based on Huawei's Pangu model. During the adjustment, Xiaoyi focused on learning the data that end consumers may generate, such as conversations, device operations, shopping, food and clothing expenses and other common sense of life. After learning to "specialize" large models, Huawei conducted word-by-word analysis and compression of prompts and output formats, ultimately halving the inference delay.

On the other hand, current quantitative techniques are also driving large models to become smaller. For example, downsize the model from FP32 to INT8. The so-called FP32 is a single-precision floating point number that can express 7 digits after the decimal point. It is quite accurate, but the operation overhead is relatively large. INT8 is an 8-bit integer, only occupies 1 byte, and the operation overhead is small. This process can be understood as fuzzifying the original delicate calculation part, and the quantitative operation allows the large model on the end side to give reasonable answers without having to achieve the "perfection" of the cloud.

After these three steps of device-cloud collaboration, hardware upgrade, and large model compression, the large model can basically be qualified and installed in the mobile phone.

But in fact, the true integration of large models with systems and applications has just begun, and a new cycle of software-driven industry is beginning to dawn.

Grab the big model and seize the next round of mobile phone sales rankings

When large models enter mobile phones, the upgrade of mobile phone software capabilities has become more important than ever.

Refer to the "spoilers" surrounding the person in charge of vivo's big model on Weibo. Among the uses of large models, vivo is divided into five categories: natural language processing, image processing, recommendation systems, artificial intelligence security, automated decision-making, and model training optimization.

Regarding mobile phone manufacturers' move to lay out large terminal-side models, Luan Jian said: "First of all, I think mobile phone manufacturers will definitely explore this possibility. If they do not explore it, they may fall behind in this field in the future."

In terms of perception on the user side, taking the image processing of image generation and image recognition and the recommendation system of product recommendation and music recommendation as examples, the addition of large models will bring a new level of intelligence to smartphones.

The former can correspond to the AI-optimized photography functions of current mobile phones. Different from the previous small program like Miaoya Camera, the photography of large models after being connected will no longer be limited to functions such as automatic P-pictures and suggested shooting locations. It can adjust the picture content in natural language and even suggest you what posture to pose for selfies. The best looking.

The latter recommendation system is expected to break the "closed" barriers of a series of APPs such as music, shopping, travel, catering, long and short videos, etc., and realize the full range of recommendation algorithms without opening the corresponding APP.

In this way, large models may change the "power division" of APPs, and it is not impossible to reshape the application software ecosystem.

More importantly, the large model will be deeply integrated with the operating system developed by mobile phone manufacturers to provide a system-level experience.

The difference in mobile phone software experience will solve the current homogeneity problem of mobile phone manufacturers "the hardware is similar and only the appearance can distinguish them."

When consumers choose smarter mobile phones, they will not need to read reviews, compare sample photos, look at running scores, look at temperature performance and other subtle differences between products. The PK of large models, mobile phone products will directly reflect the level of "intelligence level", which is one of the reasons why mobile phone manufacturers are now keen on "listing" of large models.

From AI defining cars to AI defining mobile phones and PCs, the injection of large models will reshape the sales pattern of the mobile phone market to a certain extent.

The current "dumpling" style entry of large-scale models by mobile phone manufacturers is just the beginning. Subsequent application implementation, model iteration, and experience optimization are expected to become a new round of "infinite games" for players in the mobile phone industry.

The "visible and tangible" intelligent experience from the software level will also reflect the past mobile phone conferences about "how hard-won cameras are", "performance parameters beat Apple again", and "mobile phone case materials are too extreme" The great talk of these contents injects new vitality into the content.

More importantly, large models are forcing mobile phones to increase their computing power, which will also bring about a wave of phone replacements.

In the past, due to excessive hardware configurations, there were "mobile phone users" who had used one mobile phone for five years. The upgrade experience of the large model may give them an irresistible reason to change their phones, thus bringing sales growth to the mobile phone industry.

Reprinted from 光锥智能 刘俊宏View Original

Comments

no dataCoffee time! Feel free to comment