以ChatGPT为代表的生成式大模型,展现出前所未有的通用能力,并驱动科学研究范式向科学智能(AI for science)转变。这一趋势使得高性能计算与人工智能计算加速融合,超智融合的智能超算系统成为支撑未来大模型发展与科学发现的关键基础设...以ChatGPT为代表的生成式大模型,展现出前所未有的通用能力,并驱动科学研究范式向科学智能(AI for science)转变。这一趋势使得高性能计算与人工智能计算加速融合,超智融合的智能超算系统成为支撑未来大模型发展与科学发现的关键基础设施。本文对智能超算系统在这一历史性交汇点上面临的重大机遇进行了系统分析,并深入探讨了其在计算芯片、体系结构、硬件系统、软件生态、可靠性与能耗等方面所遭遇的严峻挑战,提出需要软硬件协同设计与全产业链的紧密合作,从而为构建高效、普惠、可持续的新一代智能超算系统奠定基础。展开更多
Constructing large-scale,high-quality computational materials databases is pivotal for advancing material simulation and design.However,two essential challenges are yet to be fully resolved in this field:acquiring com...Constructing large-scale,high-quality computational materials databases is pivotal for advancing material simulation and design.However,two essential challenges are yet to be fully resolved in this field:acquiring comprehensive atomic-level structural information and effectively executing large-scale computational tasks for these material structures on supercomputers.In this study,we present a methodology that adeptly couples Artificial Intelligence(AI)and High-Performance Computing(HPC)to establish a comprehensive computational database with diverse materials data.We propose an AI-driven pipeline with a periodic-E(3)-equivariant diffusion model for structure generation and a transformer-based property prediction model incorporating 3D geometric analysis for material structure evaluation,followed by calculations on selected structures using Density Functional Theory(DFT).Specifically,a high-throughput computing framework was developed for efficient execution of various CPU/GPU/IO-intensive tasks,capitalizing on the heterogeneous computing nodes and shared storage architecture of supercomputers.Based on our HPC-AI strategy,we generated approximately 10 million hypothetical crystal structures,constituting the most extensive crystal material database currently available.By leveraging 2,000 nodes of the Tianhe-2 supercomputer for high-throughput computations,we accomplished about 80,000 DFT calculation datasets within a span of three months.Our approach represents a data-driven paradigm for boosting the materials design practice.展开更多
文摘以ChatGPT为代表的生成式大模型,展现出前所未有的通用能力,并驱动科学研究范式向科学智能(AI for science)转变。这一趋势使得高性能计算与人工智能计算加速融合,超智融合的智能超算系统成为支撑未来大模型发展与科学发现的关键基础设施。本文对智能超算系统在这一历史性交汇点上面临的重大机遇进行了系统分析,并深入探讨了其在计算芯片、体系结构、硬件系统、软件生态、可靠性与能耗等方面所遭遇的严峻挑战,提出需要软硬件协同设计与全产业链的紧密合作,从而为构建高效、普惠、可持续的新一代智能超算系统奠定基础。
基金supported by the National Key R&D Program of China(Grant No.2023YFB3002202)the Guangdong Provincial Key Area R&D Program of Guangdong Provincial(Grant No.2024B0101040005)the National Natural Science Foundation of China(Grant No.62461146204).
文摘Constructing large-scale,high-quality computational materials databases is pivotal for advancing material simulation and design.However,two essential challenges are yet to be fully resolved in this field:acquiring comprehensive atomic-level structural information and effectively executing large-scale computational tasks for these material structures on supercomputers.In this study,we present a methodology that adeptly couples Artificial Intelligence(AI)and High-Performance Computing(HPC)to establish a comprehensive computational database with diverse materials data.We propose an AI-driven pipeline with a periodic-E(3)-equivariant diffusion model for structure generation and a transformer-based property prediction model incorporating 3D geometric analysis for material structure evaluation,followed by calculations on selected structures using Density Functional Theory(DFT).Specifically,a high-throughput computing framework was developed for efficient execution of various CPU/GPU/IO-intensive tasks,capitalizing on the heterogeneous computing nodes and shared storage architecture of supercomputers.Based on our HPC-AI strategy,we generated approximately 10 million hypothetical crystal structures,constituting the most extensive crystal material database currently available.By leveraging 2,000 nodes of the Tianhe-2 supercomputer for high-throughput computations,we accomplished about 80,000 DFT calculation datasets within a span of three months.Our approach represents a data-driven paradigm for boosting the materials design practice.
