Hou Zhenyu announced that Baidu has launched a variety of AI native cloud products and is committed to reshaping large model technology in cloud computing.

At the 2023 Baidu Cloud Intelligence Conference·Intelligent Computing Conference, Baidu released new products such as the AI ​​heterogeneous computing platform "Baige 3.0", the intelligent computing network platform and the self-developed cloud native database GaiaDB 4.0 at once

Among them, Baige 3.0 has carried out comprehensive special optimization and upgrades for AI native applications and large-scale model training and inference. It supports Wanka-level ultra-large-scale AI cluster computing; compared with self-built intelligent computing infrastructure, in terms of model training and inference, the maximum throughput is increased by 30% and 60% respectively; in terms of resource utilization, Baige 3.0 can Achieving an ultra-high cluster effective training time ratio of up to 98% and an effective network bandwidth utilization rate of 95%, fully releasing the effective computing power of the cluster

In order to solve the problem of balancing the supply and demand of intelligent computing power in the AI ​​native era, the intelligent computing network platform supports global access to intelligent computing nodes such as intelligent computing centers, supercomputing centers and edge nodes built by Baidu and third parties. By connecting dispersed and heterogeneous computing resources, a unified computing network resource pool is formed. At the same time, Baidu uses the advanced computing power scheduling algorithm independently developed by Baidu to intelligently analyze the status, performance, utilization and other indicators of various computing power resources, and conduct unified scheduling. This can effectively improve the utilization rate of intelligent computing industry resources

GaiaDB 4.0 is a cloud-native database that solves the problem of single-machine computing bottlenecks by enhancing parallel query capabilities. It implements cross-machine multi-core parallel query, improving performance by more than 10 times in mixed load and real-time analysis business scenarios

Hou Zhenyu, Vice President of Baidu Group, emphasized that in the AI ​​native era, the infrastructure system for large models needs to be comprehensively reconstructed to lay a solid foundation for building a prosperous AI native ecosystem.

Hou Zhenyu said: "Large model reconstruction cloud computing is mainly reflected in three levels: AI native cloud will change the pattern of cloud computing, model as a service (MaaS) will become a new basic service, and AI native applications will spawn new research and development Paradigm.”

In terms of computing power, perform smarter calculations

In the cloud infrastructure layer, in the past, from Internet applications to mobile Internet applications, the bottom layer was based on CPU computing chips. However, as the demand for GPU or heterogeneous computing in artificial intelligence applications has increased significantly, the underlying computing power in the cloud market has begun to migrate towards GPU-based

In the third quarter of 2023, Nvidia’s revenue has exceeded Intel’s, and Nvidia’s latest market value has exceeded Intel’s US$1 trillion. In the future, the growth of GPU will far exceed that of CPU. Under this trend, we need to comprehensively rebuild the cloud computing infrastructure system for large models to support the implementation of AI native application systems

Specifically, the comprehensive reconstruction of cloud computing will be reflected in three areas, namely, the comprehensive upgrade of model-oriented intelligent computing infrastructure, data-oriented data infrastructure, and application-oriented cloud infrastructure, thus making computing Become smarter

At the model layer, large models are becoming more general, that is, Model as a Service (MaaS)

MaaS will significantly lower the threshold for the implementation of Al and achieve true inclusive benefits for Al. The new IT infrastructure it relies on will further subvert the existing cloud computing market structure at the bottom level.

According to the practical experience of Baidu Intelligent Cloud, in the past four months, since Wenxinyiyan was fully opened on August 31, Baidu Intelligent Cloud Qianfan large model platform (MaaS platform launched by Baidu Intelligent Cloud) has The number of daily API calls has increased 10 times. Customers mainly come from various industries such as the Internet, education, e-commerce, marketing, mobile phones, and automobiles. It can be clearly seen that in the past six months, many companies have really begun to use large models extensively

At the application layer, the way application development has completely changed

The unique capabilities of large-scale model understanding, generation, logic, and memory will lead a new paradigm in native application development, and the entire application technology stack, data flow, and business flow will undergo changes

In the past, CPU-based application development was mainly driven by business logic, while traditional artificial intelligence research and development required obtaining data for each independent scenario and training the model from scratch. Now, artificial intelligence native applications mainly rely on powerful large model capabilities and data-driven development. Enterprises can directly use scene data to fine-tune based on the basic large model, generate a dedicated large model, and use the model capabilities to design native artificial intelligence applications without retraining the large model. With the expansion of enterprise business, more competitive scenario data has been accumulated, thereby improving the effects of models and applications, forming a data-driven virtuous cycle

Specifically, the new paradigm of large-scale model-driven AI native application development shows several new changes:

First is the "new scene". The generative large language model has demonstrated beyond-expected capabilities in multiple dimensions such as understanding, generation, reasoning, and memory, bringing about the emergence of intelligence, which has given rise to many new business scenario applications that can be implemented, such as personal assistants. , intelligent copywriting creation, GBI (intelligent business analysis), coding assistant, etc.

The second is "new architecture". In the process of implementing these new scenarios, the large model has also produced many new system architectures, such as retrieval enhancement to generate RAG, intelligent agent, etc.

The third is "new development ecology". With large models as the core, some new tools have also appeared in the developer tool layer, including the orchestration tool LangChain, the AI ​​application development tool PromptFlow, the data framework Llamalndex, etc.

Support of data and algorithms. In terms of large models, we need to build a powerful deep learning model library and provide a variety of A native application models to meet the needs of different scenarios. In terms of intelligent computing power, we need to strengthen the computing capabilities of processors and GPUs and provide efficient computing resources to support the complex computing tasks of A's native applications. In terms of data, we need to collect a large amount of A native application data, perform data mining and analysis, and provide accurate personalized services. In terms of algorithms, we need to develop advanced machine learning algorithms to improve the intelligence level of A's native applications. Only through these supports can we build a truly prosperous A native application ecosystem

The three elements of the new paradigm of AI native application development are interdependent. Large models are the core of AI native applications, and intelligent computing provides solid support for them. The new R&D paradigm helps developers efficiently develop applications based on large model capabilities. The data flywheel is a necessary condition for successful AI native applications, allowing rapid iteration of large model capabilities to continuously improve the product experience

Hou Zhenyu said: "I believe that in 2024, truly shining AI native applications will be born."

The above is the detailed content of Hou Zhenyu announced that Baidu has launched a variety of AI native cloud products and is committed to reshaping large model technology in cloud computing.. For more information, please follow other related articles on the PHP Chinese website!