The first media of the industrial internet. Industry Family
ENIAC, the world’s first electronic computer, was programmed by six women during World War II. At that time, all women, including them, were not allowed to enter the ENIAC room. Looking back at history from the perspective of modern society, the status of women at that time presented an extreme situation—these six women’s names could not appear on the “merit list” for a long time, despite their significant contributions to modern programming.
The ENIAC, source: history.com
And this was the first human-computer interaction in history.
In the following half-century, from the “graphical user interface” revolution led by Xerox to the “touch interaction interface” revolution led by Apple, every revolution in human-computer interaction throughout history has created tremendous waves in the global PC market. This evolution not only provides users with practicality, entertainment, and convenience but also carries a greater vision of a digital “utopian” world.
Today’s “AI revolution” serves as the carrier that will make this vision a reality one day in the future. For example, current computers are merely tools, while future computers could be super assistants. They could replace users in performing basic tasks, and even creative work, by storing the user’s habits and knowledge when using the computer. More importantly, the privacy of such AI computers will no longer be an issue.
However, before this utopian “AI computer” arrives, several issues worth discussing are: How to integrate AI into personal computers? How can personal computers perform large-scale AI calculations? What requirements does this pose for hardware such as chips? Beyond hardware, what changes will occur at the software level?
As a new wave of human-computer interaction driven by AI reignites the PC market, a series of questions are becoming the focal point on the PC stage.
1
What is the quality of edge-side large models?
From smartphone cameras to smart wearables, from cars to smart homes, every input and output point can become a stage for AI to showcase its capabilities.
In early 2023, at the MWC World Mobile Communications Conference held in Barcelona, Honor launched products such as Magic OS 8.0, Magic Large Model, and Magic6 Pro phone; Qualcomm and MediaTek showcased chips with enhanced AI capabilities. That year, smartphone manufacturers such as Samsung, Vivo, OPPO, Honor, Huawei, and Xiaomi began applying large model capabilities to their smartphone products, accelerating the penetration of AI smartphones.
Thus, a revolution in AI at the terminal level has officially begun. First smartphones, then smart wearables, and next cars and smart homes, it seems that all terminals can engage with AI.
So, why are various terminal scenarios integrating with AI?
From the development trend of large model implementation, advancing from the cloud to the edge is currently one of the key development directions. If all deployment methods are cloud-based, it requires accessing the cloud through terminals, although this method can ensure sufficient parameter volume and computing power, it clearly lacks advantages for lightweight, low-latency tasks.
The core of edge-side AI lies in its ability to respond quickly to user needs, process local data, and protect user privacy without relying on cloud servers.
However, deploying large models on the edge still presents some challenges. The most obvious challenge is how to ensure that edge-side large models have sufficient computing power to support them?
From the current market’s edge-side large models, a reasonable rule is: the larger the device (the more functions), the greater the parameter volume of its edge-side large model.
In February 2024, Mianbi Intelligent, in collaboration with Tsinghua NLP Laboratory, released and open-sourced the edge-side large model Mianbi MiniCPM, with a parameter scale of 2 billion;
In September 2023, Xiaomi’s AI large model was unveiled, with a parameter scale of 1.3 billion;
In January 2023, Honor released the edge-side platform-level AI large model—Magic Large Model, with a parameter scale of 7 billion.
The aforementioned large models deployed on smartphones cannot be compared to those currently ranking in the top tier both domestically and internationally. For smart wearables with more limited functions, their edge-side large models’ parameter scale will be even smaller. However, it is worth noting that Honor’s Snapdragon 8 is equipped with a magic large model with a parameter scale of 7 billion.
In past dialogues with the media, Honor’s President of R&D Management, Deng Bin, revealed that “running large models on the edge is limited by computing power, bandwidth, and power consumption, and running a 7 billion parameter model on the edge has reached its limit.”
However, at the recent 2024 Beijing Auto Show, SenseTime’s “RiRiXin 5.0” can be described as a heavyweight bomb. It adopts a mixed expert architecture (MOE) and is the first in the country to fully benchmark and even surpass GPT-4 Turbo’s large model, with a parameter scale of 600 billion. This is a large model more suitable for deployment in vehicles.
In fact, whether it is Honor’s magic large model or SenseTime’s RiRiXin, edge-side large models share a characteristic, that is, large models with a large parameter scale deployed on the edge will certainly be compressed through “distillation” technology before being packaged to the edge.
Although large models will begin to “emerge” and present Scaling Law (scale effect) as their parameter volume increases, which is why everyone is competing for parameters; from the perspective of model application, the parameter volume of models is not always better when larger. The ideal state should be: achieve the best results with the least amount of parameters.
Therefore, large models packaged for edge deployment require “distillation” processing. For example, Google’s MobileBERT model significantly reduced the model parameter volume through knowledge distillation and other technologies, making it more suitable for deployment on mobile devices.
In addition to “distillation” processing, evolving from mobile to PC also requires chips that can ensure the normal operation of edge-side large models. In fact, over the past year and a half, many chip manufacturers have launched AI PC-compatible chips.
In October 2023, Qualcomm launched a new Arm architecture processor designed for PCs: Snapdragon X Elite;
On December 14, 2023, Intel released the Meteor Lake Core Ultra processor, which is a CPU integrated with an AI acceleration engine NPU;
In April 2024, AMD launched AI chips for commercial laptops and desktops—Ryzen PRO processors.
Not only chips, since Intel announced the “AI PC Acceleration Plan” last October, the entire industrial chain’s ecosystem is beginning to show its own shape. From independent hardware suppliers to independent software vendors (ISV), as the underlying chips and upper-level AI PCs are “in place”, some software and hardware compatibility issues will also accelerate their resolution.
However, compatibility adaptation is not an easy task. The hardware configurations of different devices vary greatly, and the diversity of operating systems and development environments requires large models to have high flexibility and portability. In addition, large models deployed on PC must also ensure user privacy, which requires solving privacy issues based on effectively utilizing local data for model adaptive learning.
2
Who are the pioneers?
“We believe that AI PCs will be a key turning point in the PC market in the coming years.” In September 2023, during Intel’s “Innovation 2023” summit in San Jose, California, its CEO Kissinger first proposed the concept of AI PCs.
This was during Intel’s large chip launch event. While showcasing its third-generation AI chip roadmap, Intel also emphasized the upcoming launch of the Core Ultra processor “Meteor Lake”, which Intel views as the most significant processor architecture change in 40 years and a chip capable of being installed on AI PCs.
This processor was officially unveiled in December 2023. In fact, as a chip giant, Intel’s progress in AI PCs has always started from the hardware level, providing underlying support for AI applications through the integration of AI accelerators and optimization of CPU architecture.
As early as 2018, Intel began its layout in AI PCs. The “Athena Project” was its attempt in AI PCs, especially in battery management, voice recognition, and security optimization. For example, Intel’s 11th generation Core processor integrates a GNA unit designed for low-power AI processing.
On the international market, Microsoft is an undeniable force. As a leader in the operating system domain, Microsoft integrated AI elements into its Windows 10 system as early as 2019, such as the Cortana intelligent assistant and machine learning-driven performance optimization features.
“Our goal is to enable every Windows PC to enjoy the efficiency boost brought by AI,” stated Microsoft CEO Satya Nadella.
In terms of AI PC layout, Microsoft focuses on the intelligence at the operating system level, continuously deepening Cortana’s functionality through Windows updates while utilizing AI to optimize system performance and user experience.
To welcome the “AI PC Year”, Microsoft also launched the Surface Pro 10 commercial version and Surface Laptop 6 commercial version in March this year.
Finally, the domestic PC manufacturer “Lenovo” is truly at the forefront of the AI PC market. In December 2023, Lenovo first released AI Ready AI PC products, including ThinkPad X1 Carbon AI and Lenovo Xiaoxin Pro 16 AI Core version. These products are equipped with Intel’s Core Ultra processor and feature a dedicated AI chip NPU to provide strong local mixed AI computing power.
“Lenovo’s AI PC will go through the AI Ready phase and enter the AI On phase in 2024, at which point everyone will have their own personal AI assistant,” said Abulike Mu Abulimi, Vice President of Lenovo Group, at the conference.
Immediately after, in April 2024, Lenovo launched new AI PC products, featuring the personalized AI intelligent agent “Lenovo Xiaotian”.
In fact, Lenovo’s vision for AI PCs is to create them as personal AI assistants, offering personalized creation, personal secretary, and device management services. The local mixed AI computing power of Lenovo’s AI PCs is one of their core features, capable of achieving multimodal natural language interaction via an embedded personal intelligent agent, as well as personal large models and local knowledge bases.
As early as 2020, Lenovo launched the Yoga series AI laptops equipped with Lenovo’s intelligent engine LCE, aimed at enhancing user experience through AI technology, such as intelligent heat dissipation adjustment, scene recognition, etc.
It can be seen that among the AI PC teams, Lenovo belongs to the “pragmatic faction”, focusing more on the practical application of AI in real scenarios, achieving intelligent optimization in heat dissipation, battery life, interaction, etc., through its intelligent engine LCE.
In addition to Lenovo, Microsoft, Intel, and other internationally renowned manufacturers, another PC manufacturer that cannot be overlooked is Apple. Just before the publication of this article, Apple held its annual release conference, and the most stunning feature was the AI-focused M4 chip.
Therefore, from the overall market structure, on one hand, brands like Lenovo, Microsoft, and Intel are leveraging their influence and technological accumulation to occupy a dominant position. Although AI PCs are still in the market cultivation period, manufacturers are exploring paths suitable for themselves.
On the other hand, emerging forces like Huawei and Xiaomi are attempting to cross over into the AI PC market, leveraging their AI experience in the smartphone field, and integrating software and hardware advantages to form a new competitive landscape.
At the model level, manufacturers are primarily providing lightweight large models for developers, such as Google, Mianbi Intelligent, and Meta.
Source: Company websites, CICC Research Department
However, in the future, with the advancement of model compression and distillation technologies, more complex large models are expected to run efficiently on PCs, such as OpenAI’s GPT series, Alibaba Cloud’s Tongyi Qianwen, Tencent’s Hunyuan, Baidu’s Wenxin series, etc., all of which could become important components of edge-side AI PCs.
In fact, the compression technology for large models has already been validated in the market. For instance, Lenovo has compressed LLM to a lightweight model for local deployment based on large model compression technology, and currently, the large model of Lenovo AI Now assistant comes from Alibaba Cloud’s Tongyi Qianwen, compressed from a 14.4GB original large model to 4GB, which can run on computers with 5-6GB of memory.
As the industrial chain matures, the AI PC era is accelerating. According to IDC’s prediction, AIPC is expected to reach an 85% penetration rate by 2027.
It is foreseeable that as AI becomes a trend of the times, this AI-driven transformation is also igniting the PC market.
3
The Future Vision of AI PCs
The term “AI PC Year” is being mentioned more and more frequently, but this track still lacks a “killer application”.
Last September, during Lenovo’s Innovation Technology Conference, while launching the “AI PC Acceleration Plan”, Kissinger called on ISVs, “Do you want to know how a ‘killer application’ is born? The answer is you are missing! It needs you to jointly shape the AI PC ecosystem and create killer applications.”
Indeed, a killer application is like a fuse that can ignite the industry for AI PCs. In the metaverse era, Apple’s Vision Pro became the underlying catalyst; and now, in the AI PC era, it relies on the entire ecosystem’s perfection.
Moreover, to integrate large models into PCs, the real issues to be solved include model compression technology, while ensuring hardware performance keeps pace, as well as ensuring privacy security and software-hardware compatibility, etc.
So, in the current wave of digital transformation, AI PCs are gradually becoming a blue ocean that technology giants are competing to layout. Industry leaders like Microsoft, Lenovo, Intel, HP, Huawei, and Xiaomi are also using forward-looking strategic vision to outline the future landscape of AI PCs.
First, Microsoft views AI as the core of future operating systems, planning to evolve Windows continuously to make AI PCs personal assistants and productivity tools for users. Currently, Microsoft is investing in advanced machine learning models for more efficient data processing, personalized services, and strengthening cloud collaboration.
Lenovo also emphasizes the idea of a “personal assistant”. Among them, building a personal AI assistant to provide personalized services is part of Lenovo’s AI PC strategic roadmap. Additionally, in terms of the ecosystem, Lenovo is also collaborating with software and hardware manufacturers through the “AI PC Acceleration Plan” to gain an advantageous position.
As a chip manufacturer, Intel is promoting the advancement of AI PCs through hardware innovation. In this regard, Intel also plans to integrate more powerful AI acceleration modules into the next-generation CPU, such as more efficient neural computing sticks, to provide low-latency, high-efficiency computing power for AI PCs. Intel’s vision is to make every PC a node for edge computing, enabling local data analysis and instant feedback while collaborating with the cloud to provide users with a seamless AI experience.
Finally, as emerging forces in the AI PC field, HP, Huawei, and Xiaomi are participating in the market transformation of AI PCs from the perspectives of AI enterprise solutions, AI chips, and smart home IoT.
Additionally, on the hardware side, to adapt to the AI PC market, upstream industry chain chips need to undergo industrial upgrades. For example, how to ensure that large models have sufficient computing power on the edge? This requires the operating system to meet the computing power needs in mobile scenarios, and battery consumption, phone storage, etc., have also been brought to the forefront.
According to Trendforce, Microsoft plans to set a minimum threshold for AIPC in Windows 12, requiring at least 40 TOPS of computing power and 16GB of memory. Moreover, structural changes have occurred on the chip side, such as architectural changes, heterogeneous computing, and memory upgrades.
With the new concepts of AI PCs endowed by international manufacturers like Intel, Lenovo, and Microsoft, the entire PC market’s explosion will follow. Whether from the perspective of enhancing productivity or from the broader context of the AI era, this round of AI transformation will surely ignite enthusiasm in the PC industry.
Looking back at the two years of “crazy growth” for large models in China, AI infrastructure has gradually matured, and PCs are the ultimate carrier for AI to land. The reason is simple; discussions about large models have already upgraded from the models themselves to AI Agents, super assistants, and intelligent agents, etc., and the best carrier for achieving this leap is the PC side.
If we look at it from the perspective of the revolution in human-computer interaction, the changes brought by AI PCs can be described as another “disruption” in the history of human-computer interaction. In the metaverse era, it was Apple that integrated AI into Vision Pro; so in today’s AI era, who will fully realize AI PCs? Who will be the ultimate winner?
“Kingdee Teng ‘Cloud’, Industry Interconnection ‘Net'”
“Tencent Adds ‘Leverage’, SaaS Opens Group Battle”
