Big model battle: what kind of intelligent computing center does cloud manufacturers need to build?
On September 13, Alibaba Cloud announced that the first batch of Tongyi Qianwen big models had been filed and officially opened to the public. Users can log on to the official Tongyi Qianwen website to experience, and enterprise users can call the Tongyi Qianwen API through Alibaba Cloud.
On September 7, at the 2023 Tencent Global Digital Ecology Conference, vps The general big language model developed by Tencent, "Hunyuan", was officially unveiled to the industry and opened to the public through Tencent Cloud. The introduction shows that Tencent's hybrid model has a parameter scale of more than 100 billion yuan, pre training corpus of more than 2 trillion tokens, strong Chinese creation ability, logical reasoning ability in complex contexts, and reliable task execution ability.
On August 31 before that, Wenxin Yiyan officially opened to the whole society. On the first day of the opening, a total of 33.42 million questions were answered by netizens. On the same day, Baidu Intelligent Cloud fully opened Wenxin Yiyan service to enterprise customers, and enterprise customers can call Wenxin Yiyan capability through Baidu Intelligent Cloud Qianfan big model platform. It is said that Baidu Wenxin's big model has iterated to version 3.5. Compared with version 3.0, the training speed has increased by 2 times, the reasoning speed has increased by 17 times, and the model effect has increased by more than 50%.
Different from other big model products, the above two big language models rely on their own cloud from training to reasoning, and their services are more diversified.
It can be seen that in the face of sluggish growth and fierce competition in the cloud computing market, cloud vendors are starting to look for new opportunities and develop new businesses from the AI model.
Slow growth of cloud market AI big model becomes a new growth opportunity
The research report released by Canalys shows that in the second quarter of 2023, the global cloud infrastructure service expenditure will increase by 16%, reaching $72.4 billion. As the market felt the impact of cost pressure, the growth rate declined compared with 19% in the previous quarter. In the last quarter, the growth rate of global cloud infrastructure service expenditure fell below 20% for the first time.
Previously, the rapid growth of cloud manufacturers was mainly attributed to large-scale enterprise digital transformation and cloud deployment. At present, the growth rate of the market is slowing down. In addition to the peak of market growth brought by the popularization of cloud, it is also closely related to the uncertainty of the global macro-economy. Cost reduction and efficiency increase have become the mainstream. Enterprises tend to adopt conservative IT strategies, and expenditures have declined significantly.
At the moment of fierce market competition and price war, cloud manufacturers need a new direction of market growth. The big model explosion brought by ChatGPT can be said to be timely. In the face of this "thousand model war", cloud manufacturers with computing power, algorithms, and data can not stand on the sidelines, and it is inevitable to enter.
Big model drives cloud vendor services and demands to meet changes
According to incomplete statistics of professional media, as of August 2023, China has released more than 130 large-scale models, and large-scale vertical models of education, medical care and other industries are also growing.
What will cloud manufacturers do in the 100 model war?
From the perspective of industry customers of cloud manufacturers, there are three main ways for them to apply the big model: first, to build a basic big model product, which requires a strong level of research and development and multiple talents, and high training costs. Only a few enterprises choose to build their own big models; Second, based on a strong basic big model, establish a big model of the industry, and iterate continuously in combination with industry data; Third, most enterprises choose to develop AI applications on the basic model and industry model.
These all contain business opportunities for cloud manufacturers, so Alibaba Cloud launched "Tongyi Qianwen", Baidu Intelligent Cloud released "Wenxin Yiyan" and ToB's big model service platform "Qianfan", and the hybrid big model was developed by Tencent but operated by Tencent Cloud. Baidu CTO Wang Haifeng believes that MaaS service (Model as a Service as a Service) will become the mainstream business model of cloud computing in the future.
On the one hand, the services provided by cloud manufacturers to customers are facing changes; On the other hand, cloud vendors also face new demands for their internal resources and capabilities.
Taking Qingyun Technology, which recently released its semi annual report, as an example, it achieved 26.51 million yuan of revenue from AI computing services in the first half of the year. AI computing services have become a very important new growth point and strategic business of the company. In the future, it will focus its core resources on this business.
Shen Shuan, president of Baidu Intelligent Cloud Business Group, said that with the arrival of the era of big models, vps The "AI acceleration chip" GPU has gradually become the center of computing power. The growth rate of computing power has far exceeded the CPU. The cloud is switching from the CPU cloud era to the GPU era.
The intelligent computing center of large-scale and cost-effective cloud manufacturers is built like this
In this case, the cloud data center previously built by cloud manufacturers is obviously difficult to meet the demand of AI computing power, and large-scale intelligent computing centers need to be built urgently.
Industry insiders said that "thousands of GPUs are often used in the training process of AI large model, and the size of the entire server cluster is often more than 100000". For cloud manufacturers, large-scale and intensive intelligent computing centers are more cost-effective in the long run, giving full play to scale benefits in resources, energy, management and other aspects, and reducing deployment, operation and maintenance costs.
This is exactly the case with Baidu Yangquan Intelligent Computing Center, which is the largest single intelligent computing center in Asia at present, with a computing power scale of 4EFLOPS (40 billion floating point calculations per second) and a total construction area of about 120000 square meters. The intelligent computing center includes 13 modules in total, which can support about 280000 servers, and uses three computing architectures: CPU+GPU+XPU.
Considering the low requirements of cost and large model training on time delay, cloud manufacturers will prefer the construction site selection of large-scale cloud data centers in the past when building intelligent computing centers, and deploy in regions with climate advantages, abundant energy, low land prices, and low electricity prices. The "East Data West Computing" western node clusters such as Zhongwei are excellent options.
Take the central defender as an example:
In terms of climate and environment, the annual average temperature of Zhongwei is 8.8 ℃, and the average wind speed is 3.3 m/s. Natural air cooling technology is suitable for the heat dissipation of the intelligent computing center, with good energy-saving effect;
In terms of power resources and electricity price, Ningxia is an important transmission channel and power energy supply point for the West East Power Transmission Project. The wind, water and fire are complementary, and the reliability of regional power supply is high, and the cost of electricity consumption is low;
In terms of land resources, most of the land where Zhongwei Data Center cluster is located is unused land on low hills and gentle slopes, with guaranteed land use indicators, and the prices of water and natural gas in Zhongwei City are relatively low.
What changes does AI model bring to cloud manufacturers and computing industry? What aspects should be paid attention to in the construction and site selection of intelligent computing center? At the upcoming 2023 Cloud Sky Conference, many industry guests gathered together to bring professional sharing and discussion.
In order to show the achievements and development prospects of Zhongwei City in computing industry, and promote exchanges and cooperation with experts and enterprise representatives in the industry, it is jointly sponsored by the Development and Reform Commission of Ningxia Hui Autonomous Region, the Department of Industry and Information Technology of the Autonomous Region, the Communications Administration of the Autonomous Region, and the People's Government of Zhongwei City, and the Data Center Committee of China Communications Industry Association Undertaken by Ningxia Zhongguancun Industrial Park Technology Investment Co., Ltd 2023 Cloud Sky Conference will be held in Zhongwei, Ningxia on September 16 。
