Public Lecture #3

"Towards Efficient AI Systems and Circuits: Progress and Perspective"

Prof. Yu WANG (Tsinghua University (THU))

Abstract

The emergence of next-generation foundation models like DeepSeek-V3/R1 has demonstrated that hardware-software co-design is important for building efficient AI systems and circuits. This talk will first analyze algorithmic evolution trends in large language models (LLMs) and methodological advances in cross-stack optimization. We then systematically explore cutting-edge developments through four key dimensions: (1) algorithmic model compression techniques, (2) large-scale distributed training systems, (3) high-efficiency inference systems, and (4) customized accelerators leveraging FPGA and near-data-processing (NDP) paradigms. The talk will integrate both latest research trends and our first-hand experimental insights.