The technological frenzy triggered by large models will eventually fall into rational prosperity.
It has been a long time since there has been such a large-scale technology boom in China's technology circles. This was fully reflected in the Alibaba Cloud Summit on April 11. In addition, Alibaba CEO Zhang Yong's Alibaba Cloud debut, the scene was crowded and could accommodate thousands of people. The hall was once closed, and both inside and outside the industry wanted to find some answers.
Zhang Yong, chairman and CEO of Alibaba Group and CEO of Alibaba Cloud Intelligence Group, said at the cloud summit that all Alibaba products will be connected to the "Tongyi Qianwen" model in the future for a comprehensive transformation. He believes that facing the AI era, all products are worth re-upgrading with large models.
In the future, every enterprise can use all the capabilities of Tongyi Qianwen on Alibaba Cloud, and can also combine the enterprise's own industry knowledge and application scenarios to train its own large enterprise model. For example, each enterprise can have its own intelligent customer service, intelligent shopping guide, intelligent voice assistant, copywriting assistant, AI designer, automatic driving model, etc.
Alibaba is not the first manufacturer in China to release large-scale models, but Alibaba Cloud, as the cloud manufacturer with the largest market share in China and the first to propose the concept of MaaS (Model as a Service), has a certain say in how to make large-scale models.
Alibaba did not focus too much on the big model itself, but revealed "cloud + AI".
Zhang Yong's Alibaba Cloud debut
This is Zhang Yong's first public appearance since he joined Alibaba Cloud, and it is also the first appearance of Alibaba Cloud Intelligence Group, which has been split into six business groups. Titanium Media App understands that the organizational changes of Alibaba Cloud are basically clear, involving the split and reorganization of various product teams, which are still in progress.
"It is not only judgment, but persistence that has brought us to where we are today." At this summit, Zhang Yong reviewed the development of Alibaba Cloud over the past 14 years, and always regarded cloud computing as the core strategy of the future. Alibaba Cloud seized the two opportunities of the mobile Internet and the industrial Internet. The investment has created today's Alibaba Cloud.
Now, when it is time for Zhang Yong to judge the future of Alibaba Cloud, he is sure that artificial intelligence is the future of cloud computing.
"Facing the future, because the full-scale explosion of artificial intelligence technology and the emergence of a new generation of technology may have a disruptive impact on all walks of life, let us see the arrival of a new era of intelligence after the digital age." He said.
Zhang Yong said that the era of digitalization can be differentiated. First, a group of new Internet companies emerged, and then all walks of life saw the power of digital technology, and everyone scrambled to start digital transformation.
But today, on the basis of the initial digitalization and transformation process, all enterprises from Alibaba to all Internet companies, from the earliest first-generation Internet companies to the latest Internet companies, and all walks of life after transformation In the face of the intelligent era, all industrial companies stand on the same starting line.
"All industries, all applications, all software, and all services are worth redoing based on new artificial intelligence technology, AIGC technical support, and large-scale model support. This will definitely bring not only an innovative customer experience, Customer service satisfaction will also bring about many changes in our production paradigm, work paradigm, and life paradigm.” Zhang Yong said.
Alibaba decided to connect all products to Tongyi Qianwen in the future and carry out a comprehensive transformation. Products such as DingTalk and Tmall Genie have become more intelligent after being connected to the Tongyi Thousand Questions Test. Like Tmall Genie, it can answer all kinds of tricky questions. After DingTalk is connected to the Tongyi Thousand Questions Test, it can automatically generate The work plan can also automatically generate a summary and to-do items after the meeting minutes, and can also take a functional sketch to automatically generate a small program.
The big model is the all-round competition of "cloud + AI"
Zhang Yong pointed out that large-scale models are an all-round competition of "AI + cloud computing". The research and development of large-scale models with over one trillion parameters is not just an algorithm problem, but includes the underlying huge computing power, network, big data, and machines. Complex systematic engineering in many fields, such as learning, requires the support of ultra-large-scale AI infrastructure.
Alibaba Cloud is the export of Alibaba’s large-scale models. In the future, Alibaba Cloud will do two things well: make computing power more inclusive, promote the positive cycle of self-developed technology and large-scale operations, and continuously release technological dividends; make AI more popular, and make everyone The products of each enterprise can be connected to the large-scale model for upgrading and transformation, and have the ability to create their own exclusive large-scale model.
Specifically, in terms of computing power and benefits, Alibaba Cloud has launched a new universal computing power type ECS cloud server Universal instance (U instance for short). The storage reserved space product OSS-RC can reduce the cost by 36%-70%. Alibaba Cloud also launched the "Feitian Free Trial Plan", which offers free trials of 50 cloud products.
In terms of the popularization of AI, Alibaba Cloud has evolved from IaaS (Infrastructure as a Service) infrastructure services to the construction of big data and machine learning platforms, big data platform services PaaS (Platform as a Services), to (Model as a Service ) The MaaS model as a service has formed a relatively complete large-scale model triple layout, which is exported to enterprise customers with general and standardized services.
In terms of IaaS, Alibaba Cloud CTO Zhou Jingren said that large-scale model training is inseparable from computing power. Today’s model training often uses the number of cards to represent the computing power behind it. In fact, this is a very simple description. All GPUs are not simple. For simple stacking, what is more needed is to allow all GPUs and heterogeneous computing power to unite to serve model training.
Alibaba Cloud has built a Lingjun cluster for AI training scenarios, which can support 100,000 cards in terms of GPU scale, connect through 800G full-speed non-congestion RDMA network, and the storage solution can achieve 20TB bandwidth per second, as well as multi-level cache.
Alibaba Cloud's machine learning platform PAI can train large-scale models with a scale of 10 trillion parameters, and can also allow thousands or even tens of thousands of cards to jointly train ultra-large-scale models.
"Because of our series of technological innovations and breakthroughs, we can support Wanka's single-task distributed training scale, the distributed acceleration ratio is close to linear, and it can have 92% linear expansion efficiency. In some training scenarios, it can bring 10x training efficiency improvement. .” Zhou Jingren said.
As the model becomes more and more complex, the cost of model reasoning and model service is also the focus of attention. In order to achieve effective model reasoning and service, Alibaba Cloud has done a lot of work in model compression, including pruning, distillation, etc. A series of compression methods reduce the cost of model services without affecting the model. On the one hand, it can reduce the delay of model service, and on the other hand, it can reduce the cost of model service to one-eighth.
Realization of "Cloud Intelligence"
The impact of artificial intelligence on cloud computing is huge. Practitioners may have the impression that in 2018, Alibaba Cloud was upgraded to Alibaba Cloud Intelligence. A senior executive of Alibaba Cloud at the time told Titanium Media App that everyone was thinking at that time, how to Driving greater cloud consumption?
"Static ones are storage, computing, and networks, while dynamic ones are databases and big data. There are also AI and IoT, among which AI is very important, which consumes a lot of computing power."
The expectations of cloud vendors are very good, but unlike computing-intensive loads, a computing task occupies a large number of computing nodes at the same time, which can easily drive a large amount of cloud resource consumption. Not good, the dispersion of AI’s own scale value makes the name of “cloud intelligence” untrue.
The large model has changed all this. The industry was pleasantly surprised to find that the large model can be integrated into almost all products and services. The computing power consumption caused by the training of the large model may be just an "appetizer". When the large model and the industry model are common enough, every day Countless reasoning is the "main course" that cloud vendors look forward to.
Just like the "Internet +" in the past, it is now "big model +". There are as many large models as there are cloud computing behind them.
Zhou Jingren said that last year, Alibaba Cloud proposed "Model-as-a-service (MaaS, Model as a Service) for the first time in the industry, which is a cloud-based and model-centric concept. Adopt this new concept to design its own cloud product system.
"The foundation of the so-called intelligent cloud comes from the intelligent infrastructure. It is necessary to be able to use big data and machine learning capabilities to the extreme. The intelligent cloud is actually the normal life extension of the cloud." Zhou Jingren added.
Cloud vendors hope that more enterprises can experience the big model, and hope that each enterprise can have an exclusive big model with its own industry capabilities based on the general big model. This will not only drive a large amount of cloud resource consumption, but also generate long-term customer stickiness , The pure resource type cannot retain customers. Once data is involved, customers tend to migrate cautiously.
"How to further reduce the cost of the exclusive model and effectively retain the part of the intelligence needed to solve industry problems, there are a lot of technological innovations and technological breakthroughs, which is why Alibaba Cloud has an advantage in this regard place," Zhou Jingren said.
Alibaba Cloud not only releases the large model, but also reveals the infrastructure behind the large model. Technology fanaticism can temporarily ignore costs and benefits, but the real prosperity of the large model cannot be separated from engineering practice and realistic commercial realization. .
(This article is the first author of Titanium Media APP｜Zhang Shuai, welcome to add the author’s WeChat mr3right to communicate)