Baidu Smart Cloud recently successfully launched the country's first self-developed Kunlun Chip third-generation Wanka cluster. This milestone breakthrough not only marks an important step for Baidu in the field of artificial intelligence computing power but also provides new development ideas for the entire industry. With continuous advancements in technology, the enhancement of computing power is crucial for supporting the training and application of large-scale models.
Over the past year, as AI technology has become more widespread, many companies have faced tight computing power issues, which directly led to high costs for using large models. Baidu stated that through the development of self-researched chips and the construction of the Wanka cluster, they have not only effectively solved their own computing power supply problems but also provided a reference and support for other enterprises. The Wanka cluster has ultra-large-scale parallel computing capabilities, significantly improving training efficiency, especially when training complex models with hundreds of billions of parameters, which can greatly shorten the training cycle.
Image Source Note: Image generated by AI, image licensed by Midjourney
The application of the Wanka cluster will meet the demand for rapid iteration of AI-native applications and can also support the processing of trillion-parameter models and multimodal data, providing strong support for the development of Sora-like applications. Additionally, the multi-task concurrency capability of the Wanka cluster allows it to dynamically allocate resources to train multiple lightweight models simultaneously, achieving efficient use of computing power. This innovation by Baidu Smart Cloud not only enhances the overall utilization rate of the cluster but also significantly reduces the cost per unit of computing power.
However, issues such as past multi-chip mixed training and increased failure rates have become major challenges in the deployment of the Wanka cluster. To address these problems, Baidu launched the upgraded version of the Baibei AI heterogeneous computing platform 4.0 in September 2024, which plays a crucial role in the construction of the Wanka cluster. Through model optimization, parallel strategies, and dynamic resource allocation, Baidu Smart Cloud is promoting the effective utilization of computing power, laying the foundation for future AI applications.
The success of Baidu Smart Cloud not only demonstrates the strength of independent research and development but also injects new momentum into the vigorous development of domestic large models. In the future, with the continuous expansion and optimization of the Wanka cluster, we look forward to more innovative AI applications being implemented, bringing new opportunities for industry development.
