During AI/ML training processes, storage is crucial for enhancing the overall training efficiency, especially after the training tasks and datasets have grown substantially in scale. How to accelerate the flow of data among different tasks and GPU devices becomes even more important. The speed of data loading directly determines the efficiency of model training.

Based on these premises and analyses, we believe that local storage can play a more significant role in AI/ML training tasks, particularly in the acceleration of the link between data and GPUs. Remote storage and local storage should coordinate to leverage their respective advantages, achieving the maximization of storage efficiency.