“巢湖明月”合肥新一代算力集群-合肥人工智能计算中心智算资源

合肥人工智能计算中心建设规模为智能算力100PFlops@FP16，数据存储容量2PB。中心采用国产昇腾技术路线，通过搭载448颗国产AI芯片昇腾910，提供99PFlops@FP16的人工智能算力，支撑人工智能训练、加速模型训练过程、提升模型训练精度，加快人工智能算法和应用创新。另外，1PFlops@FP16的算力支撑人工智能应用测试和验证。借助ModelArts人工智能使能平台，资源得到了统一管理、分配和调度。该平台提供先进的AI引擎、开发环境、模型训练以及模板管理和部署功能。合肥人工智能计算中心从芯片到系统再到顶层应用部署，形成了一个端到端的全栈解决方案，并且全部实现了国产化，确保了自主可控性。The Hefei Artificial Intelligence Computing Center has a designed computing scale of 100 PFlops@FP16 intelligent computing power and a data storage capacity of 2 PB. Adopting the domestic Ascend technology roadmap, the center is equipped with 448 domestic Ascend 910 AI chips, delivering 99 PFlops@FP16 of AI computing power. This capacity supports AI training, accelerates the model training pipeline, enhances model training accuracy, and promotes the innovation of AI algorithms and applications. In addition, the remaining 1 PFlops@FP16 computing power is used for AI application testing and verification. With the support of the ModelArts AI enabling platform, unified management, allocation and scheduling of computing resources are implemented. This platform offers advanced AI engines, development environments, model training capabilities, as well as template management and deployment functions. The Hefei Artificial Intelligence Computing Center has built an end-to-end full-stack solution covering from chips, full systems to top-level application deployment, which is fully domestically localized to ensure self-reliant and controllable computing capabilities.

搜集汇总

背景与挑战

背景概述

该数据集提供合肥人工智能计算中心的智算资源，采用国产昇腾技术路线，提供99PFlops@FP16的人工智能算力和2PB的数据存储容量，支持人工智能训练和应用创新。中心还提供多个平台服务，促进政产学研用的融合。