Qiheng Lü
Vice President of Moore Threads and Dean of Moore Academy
Qiheng Lü is Vice President of Moore Threads and Dean of Moore Academy, as well as a Senior Engineer. He has extensive experience in domestic GPU ecosystems, technology strategy, and industry-academia collaboration. His work focuses on cutting-edge fields such as large language models, heterogeneous computing, AI for Science (AI4S), and Physical AI. He is committed to fostering collaboration among chip manufacturers, software developers, developer communities, research institutions, and industry users. He is responsible for technology planning for Moore Threads' embodied AI products.
Topic
The Computing Foundation and Ecosystem for Physical AI: From Infrastructure to Real-World Deployment
Embodied intelligence is moving beyond isolated model experiments in the lab toward a complete system encompassing data collection, simulation and synthetic data generation, model training, post-training, evaluation, cloud or workstation inference, and deployment on physical platforms. The core challenge of Physical AI lies not simply in scaling up models, but in the new combination of workloads involving vision, language, action, simulation, and real-world interaction. These workloads place distinct demands on GPU compute, memory capacity and bandwidth, interconnects, software stacks, toolchains, and developer ecosystems compared with traditional large language models. Starting from the characteristics of embodied AI workloads, this talk will explore how to build GPU systems and ecosystems tailored for Physical AI, and identify the engineering capabilities and collaborative efforts needed to move from technical validation to real-world industrial deployment. Audience Takeaways: 1. Develop an end-to-end technical understanding of Physical AI, from data collection and training to simulation and deployment. 2. Understand the key differences in compute requirements between embodied AI workloads and traditional LLM inference. 3. Gain a practical framework for evaluating the maturity of GPU platforms, software stacks, and developer ecosystems for Physical AI.
