Afraid of "scrapping and returning"... Is the "Hundred Model Battle" that burns money interrogating computing power inclusive? Domestic chip requests to enter the market: hoping to "use" giants | unable to afford | Nvidia | large models | computing power | Zhao Lidong | domestic chip | Suiyuan Technology
The big model is too appealing. An engineer asked ChatGPT to write a love letter in the tone of science fiction writer Liu Cixin, but the wife received it and tears streamed down her face. Help radiologists to write professional year-end summaries, and provide the creators with poems combining Tianjin Kuaiban and RAP rap style, which are of great significance to the big model.
The demand for AI computing power has changed the global semiconductor industry landscape. On July 9, 2020, which could be a milestone day in the semiconductor industry, Nvidia, with AI computing power as its strength, surpassed the semiconductor industry leader Intel for the first time in terms of market value. Nowadays, Nvidia's market value has jumped to the trillion dollar mark. Such a remarkable performance is enough to be recorded in the history of Wall Street.
Industry forecast data shows that by 2030, AI high-performance computing will account for 40% of the global semiconductor industry's nearly trillion dollar market. This means that behind the current frenzy of big models, there is a cold thinking about "muscle display" that is difficult to avoid. The widely recognized statement in the industry is that the 10000 Zhang Yingweida A100 is the basic computing power threshold for a good large model.
Based on this, the ongoing "Hundred Model War" in China must face the soul torture of "not being able to afford it" and bring up the topic of "inclusive computing power". On July 6, 2023, at the chip themed forum of the 2023 World Artificial Intelligence Conference "From End to Cloud, Bravely Climbing the Peak of Chip", the AI industry felt both pressure and motivation - large models brought massive computing power demand, and also brought new opportunities for domestic chips to counterattack with "innovative architecture+open source ecosystem".
Academician of the Chinese Academy of Sciences and President of Shenzhen University, Mao Jun, delivered a keynote speech, pointing out that "chips are just means, and microelectronic systems are the purpose." The forum is hosted by the Office of the Organizing Committee of the World Artificial Intelligence Conference and hosted by the Shanghai Integrated Circuit Industry Association.
Is computing power inclusive a bit far away?
At this year's World Artificial Intelligence Conference, more than 30 large models competed to attract attention, including general models such as Huawei Pangu, Alibaba Tongyi, iFlytek Spark, Baidu Wenxin, Fudan Moss, dialogue models such as Shangtang Shangliang and Yunzhisheng Shanhai, music models such as Tencent X music, and multiple vertical industry professional models. Just recently, the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology also released the "Research Report on China's Artificial Intelligence Large Model Map", pointing out that China has released 79 large models with parameter scales of over 1 billion, and the "Hundred Model Battle" is imminent.
Zhao Lidong, founder, chairman, and CEO of Suiyuan Technology, objectively analyzed the computational power requirements of these large models for exponential growth in the future——
In the embryonic stage of large model technology, based on previous experience, training a large model with GPT3 parameter scale requires 10000 Zhang Yingweida V100, which takes 14.8 days, and the training cost is extremely high. Later in the second half of this year, China's large-scale models are expected to enter a period of technological expansion and application germination, starting the stage of inference deployment. Based on Google's experience, replacing 320000 queries per second with a large model would require deploying approximately 4.1 million Zhang Yingweida A100, resulting in an additional inference cost of approximately $36 billion. Afterwards, as we move forward to the period of accelerated application expansion, the application of large models truly empowers thousands of industries, and the required computing power comes with a wave of multiplier effects. For example, WeChat has a monthly activity of billions and nearly ten million mini program ecosystems, and its computing power demand is simply astronomical.
No matter how fanatical the training and reasoning of large models may be, they will eventually return to business logic, and achieving economic benefits is fundamental, rather than "ignoring the cost". This involves whether computing power can be as universal as water and electricity. The so-called "Pu" refers to being able to obtain it when needed; The so-called "benefit" refers to being able to afford it. But at the chip themed forum, attending experts have to admit that China still has a long way to go before computing power becomes universal.
Is there a "window period" for catching up and surpassing?
How to achieve inclusive computing power? Mao Junfa and Zhao Lidong, academicians of the CAS Member, both hold the view that innovative architecture plus open source ecology.
According to the reporter's understanding, the current AI chips in the semiconductor market are mostly based on Nvidia's GPU and CUDA as the mainstream architectures, and many applications are also developed based on these mainstream architectures. But in the long run, the domestic chip industry must establish a second solution to provide the market and customers with a second choice.
In fact, computing power is so expensive or even impossible to obtain, which is undoubtedly an excellent time window for domestic chips to strive for independent innovation architecture and meet the diverse demands of the market for cost-effectiveness and energy efficiency. "The convergence in the number of large models and the concentration and intensification of the ecosystem are very favorable opportunities for establishing a new ecosystem of AI chips," said Zhang Yalin, founder and COO of Suiyuan Technology.
On the forum, calls for the path of open source sharing have also been made from time to time. Experts believe that for commercial reasons, some vendors may not choose open source, which is understandable. But if you want to find a way to catch up, you cannot bind yourself and set limits on technology. Zhang Suxin, President of the Shanghai Integrated Circuit Industry Association, called on the forum that a lone tree does not make a forest. Domestic chips should gather together, expand the participation of developers, form an ecological loop, and accelerate industrial development and even internationalization.
Can domestic chips play more games?
Various encouraging phenomena indicate that some domestically produced chips are accelerating their inspirational journey.
For example, Suiyuan Technology focuses on artificial intelligence cloud edge computing technology products, based on self-developed chip architecture and programming framework, independent of foreign manufacturers or ecosystems, and independently creates two product lines for cloud training and inference. It has a complete solution covering chips, boards, systems, clusters, and computing platforms. Suiyuan Technology, founded on the eve of the first World Artificial Intelligence Conference, is not yet 6 years old, but has entered a positive cycle of development from independent research and development to mass production, from commercial application to optimization and iteration.
The 1280 card scale fully liquid cooled AI training cluster of Suiyuan, which has been established in large-scale laboratories, is a computing power cluster tailored for large-scale AI computing infrastructure in China, with a core carbon reduction index PUE of 1.1 or less.
The reporter found that many chip manufacturers are unwilling to follow in the footsteps of giants and also crave the support of the market and government departments. The development of the chip industry has a rhythm and stages, and the government may explore providing segmented support. In the stages of pre training, fine tuning, and inference, the support will gradually increase to better promote the development of the AI chip industry.
"With more appearances, the results of the competition naturally improve. No one can directly step onto the podium from the audience." Zhao Lidong said, "Chips are 'used in, discarded out', the more they are used, the better they can be used. In a gradual process, we must cultivate a computing ecosystem and iterate computing products. This process is something we must go through."
The China Electronics Technology Standardization Research Institute of the Ministry of Industry and Information Technology has collaborated with the National Key Laboratory of Communication Content Cognition of the People's Daily to establish an overall evaluation and evaluation system that includes performance evaluation, scenario evaluation, and comprehensive evaluation of artificial intelligence chips. In order to jointly promote the high-quality and orderly development of China's artificial intelligence chip industry, the two sides have jointly led the "Smart Yue Plan".
