Intelligent computing centers are built in large quantities, and experts warn against "AI overheating"
Author/ IT Times reporter Mao Yu
Editor/ Hao Junhui Sun Yan
After the Spring Festival in 2024, China's computing power market is undergoing deep transformation and structural adjustment.
Due to the high investment and financing pressure, many startups have turned to vertical tracks with more potential and targeting, such as government models, autonomous driving, medical radiation, etc., because they cannot afford the huge overhead required for large model training.
"At present, the market is still overheating AI, and many small and medium-sized manufacturers have fallen in front of the huge investment costs of general models, and we need to be wary of the overcapacity of the photovoltaic industry a few years ago, and we should seek a new breakthrough this year. Peng Lu, COO of Shanhai Engine, told the "IT Times" reporter.
GPU, the base of computing power, is still the key word in the current computing market.
In the second-hand market, the price of computing power of Nvidia series chips has declined significantly recently, down about 10% from before the Spring Festival, showing changes in market demand. Since April, the Flush Computing Power Leasing Concept Index has continued to fall, falling by more than 16% as of April 23.
An industry insider told the "IT Times" reporter that "after the construction of this wave of intelligent computing center, the market competition has basically been sufficient, and the price of computing power will not be too outrageous." ”
Small computing power operators can't play anymore
Although large technology companies, such as Baidu's Wenxin Yiyan and Alibaba's Tongyi Qianwen, still maintain high investment in the field of large models, the market generally believes that as the cost of computing power soars, only a very small number of companies with strong financial strength can support such huge expenses. In addition, behind the high financing of domestic companies such as MiniMax and Dark Side of the Moon, it reflects the intensified competition in the industry and the pressure of business model exploration.
According to public information, MiniMax and the dark side of the moon are all owned by Ali. The founders of the former, Yan Junjie and Zhou Yucong, are both from SenseTime and are also known as "SenseTime" startups, while the latter has become a unicorn with the highest valuation among domestic large-scale model startups after completing its Series B financing in February this year.
However, limited by the U.S. chip control policy, the special version of the H20 chip launched by NVIDIA for the Chinese market has not met market expectations in terms of cost performance and applicability, and the procurement needs to be strictly reviewed, and the domestic shipments have been affected to a certain extent, and most of the Nvidia chips on sale on the market are second-hand chips.
On an e-commerce platform, a reporter from the "IT Times" searched and found that many service providers are still selling complete machine services, and a server manufacturer sells server hosts equipped with 8 cards H100 and A100, with prices ranging from 1 million ~ 3 million yuan. After consulting customer service, the reporter learned that the current price is still preferential to negotiate, and later the reporter got a more detailed quotation, which was reduced by 200,000 ~ 600,000 compared with the homepage display price, and promised to enjoy a three-year warranty for the merchant.
On a second-hand platform and social software, there are also sellers who sell Nvidia or Huawei Ascend constantly online "shouting", with the text saying that "the low price is negotiable" and "the quotation is valid for one week, subject to the latest quotation".
"The warranty mentioned by the merchant above is the merchant warranty, not the official warranty of NVIDIA, so the price is easy to negotiate. The above-mentioned industry insiders said.
"At present, large factories are still betting heavily on large models, and they have funds and are willing to invest. But as far as I know, many startups have withdrawn from the large-scale model track. Many computing power operators are facing survival challenges due to high hardware investment and maintenance costs, and cannot afford to purchase large-scale computing power or help customers deploy it on a large scale. Peng Lu believes that it is expected that by the end of 2024, a number of small and medium-sized players will exit the domestic computing power track.
The price of computing power stabilized
At present, the domestic computing power market is changing from last year's seller's market to a buying and selling balance, and the upstream and downstream of the industrial chain are seeking a development model that adapts to the new market environment, and the selling price of computing power is also stabilizing.
Not long ago, a platform called "Supercomputing Internet" played 8*A800 80GB server, with a promotional price of 2 yuan per card per hour, compared with the market price of 8~10 yuan, which hit a "fracture".
"This price is still a gimmick, and the platform only takes out a total of 100,000 yuan, and at most 9 8-card servers are used for a month. An industry insider told the IT Times that this price is not universal.
However, compared with the "jumping" price increase in the market at the end of last year, after entering the spring of March, the price of computing power has indeed decreased. In addition, since most of the large Internet companies have built their own computing centers, IDC service providers that have hastily launched intelligent computing centers have also begun to adjust their strategies, from simply providing bare metal or complete cloud server leasing services to providing more value-added services and PaSS services.
"Many traditional enterprises don't know what AI can do for themselves, in fact, the real demand for computing power has not been stimulated. Yu Teng, the founder of Sober Heterogeneity, told the "IT Times" reporter that what is urgently needed in China now is to open up the industrial chain of large model landing, so that traditional business owners can truly see the value of AI and computing power.
Lv Tianwen, vice chairman of the China Electronic Energy Conservation Technology Association, said in an interview with the "IT Times" reporter at the beginning of the year, "Two years later, the domestic computing power market will tend to stabilize, and elimination and transformation are the outcomes of some companies." But this is a good thing, the thousand model war and the ban on sales policy have spurred the progress of domestic computing power, and small and medium-sized manufacturers can also find the right position as soon as possible. ”
Autonomous driving and medical care lead the vertical track
China's large-scale model companies are exploring deep cultivation of subdivisions in order to make breakthroughs. Liu Liehong, director of the National Data Bureau, revealed at the 2024 annual meeting of the China Development Forum held in March that the number of large models with more than 1 billion parameters in China has exceeded 100, and the industry large models have deeply empowered electronic information, medical care, transportation and other fields, forming hundreds of application models and empowering thousands of industries.
"At present, we have received more demand for autonomous driving, followed by the medical track, and technical departments such as scientific research institutions and medical imaging are very suitable for the use of intelligent computing. Peng Lu revealed.
At present, among the new energy vehicles, led by Huawei, Xiaomi, Ideal, Geely and other car companies are all increasing the layout of large models. Previously, JD Health officially released the large model for the medical and health industry "Jingyi Qianxun", Baidu released the first domestic "industrial-level" medical model "Lingyi Model", and the Medical Federation officially released the self-developed medical large language model MedGPT, etc., all of which showed development prospects beyond market expectations.
Data labeling is an important bottleneck
From catching up with OpenAI to refining vertical applications, China's computing power market is currently entering a critical stage.
"On the vertical track, China's large model has a certain competitive advantage, because the United States is limited by local power consumption, legal supervision and other issues, and large-scale deployment in some regions may face difficulties. The above-mentioned industry insiders said.
However, with the gradual rise of the vertical track of large models in China, data labeling has become an important bottleneck restricting the development of the industry. For the data annotation of speech and image recognition in traditional artificial intelligence, it relies on absolutely accurate simple annotation, which does not require annotators to master professional domain knowledge, but only needs to have ordinary human general knowledge. However, the training effect of vertical large models depends on accurate data and reinforcement learning (RLHF) from human feedback. That is, there is a need to guide the "evolution" of intelligent systems by using human feedback on the results.
Source: Oriental IC
Data annotators who provide feedback need to be highly professional, especially with industry knowledge. Especially in the fields of strong professionalism and highly accurate annotation, such as architectural design, aviation simulation, etc., the relevant data processing cost is high and talents are scarce, and the development of the field is affected to a certain extent.
It is necessary to accelerate the formation of a national integrated computing power system
In 2024, generative AI will enter an acceleration period. Open AI has continuously released multi-modal large model "explosions", and the United States has introduced a new policy on export controls on computing chips. Problems such as insufficient number of computing equipment, insufficient independent research and development capabilities, and high costs are restricting the development of the current domestic general large model.
The above-mentioned people believe that compared with GPT-4 and the upcoming GPT-5, China's large model still has a lot of room for improvement, and Chinese companies are still facing severe challenges in global competition.
The Intelligent Computing Center is becoming a new driving force to support the rapid development and economic growth of the artificial intelligence industry. The Action Plan for the High-Quality Development of Computing Infrastructure proposes that by 2025, the scale of computing power will exceed 300EFlops, and the proportion of intelligent computing power will reach 35%.
The "Guidelines for the Innovation and Development of Intelligent Computing Centers" released by the State Information Center predicts that during the "14th Five-Year Plan" period, when the intelligent computing center achieves 80% application level, the investment of cities (regions) in the intelligent computing center can drive the growth of the core artificial intelligence industry by about 2.9 to 3.4 times.
Time is running out, this "national strength race" is looking forward to a more powerful engine, and at the same time, we must also beware of the waste of resources caused by the construction of "swarms" in various places. According to incomplete statistics, more than 30 cities across the country have invested in or are preparing to invest in the construction of intelligent computing centers, and "who is the computing power sold to" is a mandatory question for these intelligent computing centers.
During the two sessions this year, the mainland's "Government Work Report" proposed for the first time to build digital infrastructure in advance, accelerate the formation of a national integrated computing power system, and cultivate a computing industry ecology.
Typesetting / Ji Jiaying
Image / Supercomputer Internet Oriental IC
来源/《it时报》公众号vitimes
E N D
Please add a "star" so you don't miss us