面向AI训练的数据存储技术研究OA

The development of data-driven artificial intelligence(AI)is advancing rapidly,with the underlying layer performing training and optimizing algorithms through large amounts of data.The parameter scale of AI large models has reached billions,evolving from single modal to multimodal,involving types such as text,language,and video,and the data scale has grown from tens of terabytes to petabytes.The increasing size of model parameters means that the efficiency requirements for training become higher,and improving storage performance becomes an important direction to enhance the training efficiency of large models.The paper analyzes the storage access requirements of different stages of the large models,as well as the matching sit-uation of different distributed storage for the training of the large models,and explores and proposes directions for improving the performance of distributed storage.