Yesterday, the first batch of terminal products equipped with Intel Core Ultra processors were launched for sale in China.
Gao Yu, general manager of technology of Intel China, introduced at the press conference that Intel Core Ultra processor can be competent for up to 20 billion parameter big language models. At present, it has carried out adaptation cooperation with Baichuan Intelligent, ERNIE Bot, Zhipu AI, WPS AI, Tongyi Qianwen, OPPO AndesGPT and other big models in China.
IDC predicts that by 2024, more than 70% of AI enabled terminal devices will be installed in the Chinese market, with AI terminals accounting for 55%. The proportion of new machines assembled by AI PCs in the Chinese PC market will rapidly increase in the coming years, reaching 81% by 2027, becoming the mainstream of the PC market.
Gao Yu expressed in an interview after the meeting that he is optimistic about the rapid development of NPU based application scenarios and AI mapping applications. However, he also admitted that the AI PC ecosystem has just begun construction and still needs a long time to improve the industrial chain.
NPU based application scenarios will develop the fastest
Currently, both Intel and AMD have launched processor solutions with built-in NPU units, and Qualcomm has also released built-in NPU processors based on ARM architecture.
Many industry analysts have pointed out that the collaborative architecture of "CPU+NPU+GPU" will become the future AI PC computing power base. Among them, NPU is suitable for continuous AI and sharing AI load, while CPU is suitable for low latency AI tasks with single inference.
Gao Yu stated that compared to calling cloud based models, the advantage of AI PCs is that they do not rely on networks and are highly customized for personal scenarios, providing AI services for thousands of people and faces.
Huang Yiqun, General Manager of Sales Department of Microsoft Asia Pacific Equipment Partner, introduced the effect of Windows Meeting Room software after applying NPU: online meetings can achieve automatic framing, including blurring the background of the meeting and automatically positioning the portrait in the middle of the screen.
The XPU architecture is also expected to rapidly popularize AI digital humans. According to Gao Yusheng, through the division of labor and cooperation among NPU, CPU, and GPU, NPU is responsible for human motion capture, CPU is responsible for logical calculation, and GPU is responsible for 3D rendering, quickly generating digital humans from real people on PC.
In the live demonstration, Beijing Academy of Painting used AI PC to create real-time Qi Baishi style paintings in the form of illustrations.
In addition, Intel is also collaborating with Jianying to localize and deploy the AIGC model. In the on-site demonstration, Jianying optimized the "intelligent image cutout" function through the Core Ultra platform, which can reduce the time consumption by up to 2/3.
When it comes to the prospect of AI PC application scenarios in 2024, Gao Yu believes that NPU based application scenarios will develop the fastest. "For example, in Jianying, using NPU to cutout images directly reduces power consumption by a large proportion and shortens time by one-third. This type of application scenario combined with NPU is definitely the fastest to land."
Secondly, AI mapping applications similar to Stable Diffusion will also have rapid development. "We can now achieve fast image generation in 1 second and high-definition image generation in 6 seconds on a PC. This speed is completely on par with cloud computing. I believe that the development of text-based and image-based graphics will be faster. However, large-scale language models for pure text interaction still need to be explored, and may be slightly slower."
Compared to the booming development of mobile applications, the development of PCs in the application layer has been relatively slow in the past few years. However, Gao Yu believes that AI PCs will change this situation.
"The applications on Windows in China are not as diverse as those on mobile phones, with a difference of two in quantity. Nowadays, AI PCs have created demand and lowered development barriers to some extent. Therefore, many small software vendors are expected to combine AI with PCs to create new things. At this time, it may be necessary to use app stores to release applications, which is likely to appear in the near future."
The AI PC ecosystem is still under construction
Although a hundred model war has begun in China, there are still too few AI native applications developed on top of the models. Gao Yu believes that this is due to the ongoing construction of the industrial chain ecosystem.
"There won't be so many early AI applications, but let's not be limited in our thinking. AI is changing very quickly. A year ago, we couldn't even imagine it was like this today, and two years ago, we couldn't even think of it. So, can we now think of what will happen in one or two years? But the predictable trend is that the industry is moving towards this direction."
At the press conference, Intel announced its hardware and channel partners for the Core Ultra processor, including Asus, Acer Acer, Dell, H3C, Honor, HP, Lenovo, Xiaomi, JD.com, Pinduoduo, Suning.com, Tmall, etc; Software and model partners include Microsoft, Kirin Software, Tencent Meeting, NetEase Youdao, iQiyi, Jianying, etc.
Gao Yu admitted that the construction of the AI PC ecosystem has just begun and still takes a long time, but once the ecosystem is established, the situation at that time may be completely unimaginable.
Nowadays, everyone is still exploring. This concept has not been born yet and is still evolving rapidly. The first thing we need to do is to adapt to large models, including language class models and textual graph models. This is the most essential ability of AI PCs, so that they can run smoothly and with minimal accuracy loss on the PC, and also strive to minimize memory usage and GPU usage. There is a lot of optimization work involved in this. After solving this problem, we will proceed with the following steps: The industry will iterate rapidly, giving birth to many applications that are unimaginable today