As Time-Series Foundation Models (TSFMs) like Timer rapidly scale up, a critical gap has emerged between algorithmic capabilities and underlying system infrastructure. In massive Industrial IoT scenarios, traditional data pipelines (e.g., exporting databases to CSV/Numpy) introduce severe I/O bottlenecks. Furthermore, fine-grained MoE time-series architectures face immense memory and communication walls during training. This session introduces a high-performance, end-to-end AI training infrastructure built natively on Apache TsFile.

Speakers:

Xinhao Gu: Ph.D. Student, Tsinghua University | Apache IoTDB Committer | Apache TsFile Committer

Xinhao Gu is a Ph.D. student at the School of Software, Tsinghua University, and an active contributor to the open-source community, serving as an Apache Committer and the core member of AINode in Apache IoTDB. His research primarily focuses on AI Infra and MLSys, with a specific emphasis on the efficient training and system integration of Time-Series Foundation Models. He is dedicated to solving end-to-end system bottlenecks—from massive industrial time-series data storage to distributed deep learning frameworks—pushing the evolution of time-series infrastructure into the AI era.