On September 25, Baidu AI Cloud announced at its 2024 Baidu Cloud Smart Conference that it would comprehensively upgrade the two AI infrastructures of Baige AI heterogeneous computing platform 4.0 and Qianfan big model platform 3.0, and upgrade the three AI native application products of code assistant, intelligent customer service and digital human, respectively, for computing power, models and AI applications.
"In 2024, the industrial landing of the big model is accelerating. At present, on the Qianfan big model platform, Wenxin's big model has been adjusted more than 700 million times a day, helping users fine tune 30000 big models and develop more than 700000 enterprise level applications. In the past year, Wenxin's flagship big model has reduced the price by more than 90%, and the main model is free of charge, minimizing the cost of enterprise innovation trial and error." Shen Shao, executive vice president of Baidu Group and president of Baidu AI Cloud Business Group, said that the big model and the supporting computing management platform, model and application development platform are rapidly becoming a new infrastructure.
Big computing power is the fundamental condition for the implementation of large models. It is reported that in order to meet the full journey computing needs of enterprise landing large-scale models from cluster creation and development experiments to model training and reasoning, and to address the two challenges of high cost and difficult operation of large-scale GPU clusters, Baidu AI Cloud upgraded and released the 100Ge AI heterogeneous computing platform 4.0, and comprehensively upgraded computing management capabilities for 10000 card and 100000 card clusters.
During the cluster creation phase, enterprises typically need to perform a large amount of complex and tedious computing power configuration and debugging work. Baige 4.0 comes pre installed with mainstream large model training tools, which can achieve tool level deployment in seconds and reduce the preparation time for running a 10000 card cluster from a few weeks to 1 hour, greatly improving deployment efficiency and shortening business deployment cycles.
In the development experiment stage, enterprises need to conduct multiple tests on models with different architectures and parameters based on business goals, and then develop the best model training strategy to ensure the performance and effectiveness of subsequent training. The newly upgraded observability system of Baige 4.0 can comprehensively monitor aspects such as multi-core adaptation, cluster efficiency, and task automatic fault tolerance, providing intuitive decision-making basis and helping users better control the overall project.
At present, Baige has achieved an effective training time ratio of over 99.5% on the Wanka cluster, leading the industry and greatly saving customers' computing power and time costs. In addition, Baige 4.0 has significantly improved the model training efficiency of the cluster through a series of innovations in cluster design, task scheduling, parallel strategy, and video memory optimization, with an overall performance improvement of 30% compared to the industry average.
In addition, in order to meet the needs of enterprise customers for model invocation, model development and application development, Baidu AI Cloud released Qianfan Model Platform 3.0. In terms of model calling, the upgraded Qianfan platform can not only call nearly a hundred domestic and foreign large models, including the Wenxin series, but also support calling various traditional small models such as voice and visual. While expanding the types of models, Baidu AI Cloud continues to reduce the cost of model invocation.
The continuous improvement of tool platforms has also promoted the explosive growth of the large-scale model industry in the past year. The Qianfan platform has precipitated eight industry solutions, including manufacturing, energy, transportation, government affairs, finance, automobile, education, and the Internet.