幸福魔方

凤凰网科技讯 8月24日，摩尔线程正式发布《MTT S5000 Prefill-as-a-Service技术白皮书》，基于旗舰级AI训推一体智算卡MTT S5000，提出Prefill-Decode异构解耦方案，面向AI Agent、代码生成、超长文档分析等长上下文推理场景，旨在破解推理成本瓶颈。白皮书指出，大模型输入上下文规模进入百K至1M级别后，Prefill（预填充，计算密集型）与Decode（解码，访存密集型）在同一资源池混跑造成结构性算力错配。         A man in East China's Jiangsu Province who recently tried to swim across the Yangtze River to Shanghai to save travel costs is rescued by police. Photo: Screenshot from Jiangsu TVA man in East China's Jiangsu Province who recently tried to swim across the Yangtze River to Shanghai to save travel costs was rescued by police after drifting for nearly two hours in the busy shipping channel, exhausted and stricken with leg cramps, Jiangsu TV reported.
Police received a report around noon on August 4 that a man was swimming near a park, where heavy vessel traffic made the situation particularly dangerous. After a coordinated search by land and water, officers located the man, according to multiple media reports.
However, the man initially refused to board the police boat. Officers then learned that he intended to swim across the Yangtze River to Shanghai to avoid paying for transportation, according to Jiangsu TV.
The section of the Yangtze in Nantong is wide and has strong currents, while the main shipping channel is busy with passing vessels. The attempt posed serious risks not only to the man himself but also to the safety of vessels navigating the river, policemen said, per media reports.
After nearly 30 minutes of persuasion, the man finally abandoned his plan and boarded the police boat. He had by then spent about two hours in the water and was exhausted, suffering from severe leg cramps. Police officers provided him with food and water and took him to a police station to rest.
The police later identified the man as a native of Linyi, East China's Shandong Province, and found that he had a mental health condition. They contacted his family and the Nantong rescue center and arranged for him to undergo a psychiatric assessment.
His family subsequently took him back to Shandong and thanked the officers for rescuing him. 
Global Times
。解决方案是将二者彻底解耦，分别运行在各自最契合的硬件资源池上，实现单位Token基础设施成本大幅下降。MTT S5000硬件特性与Prefill任务高度契合：提供高稠密算力压缩首字时延，原生支持FP8全精度计算，并全面兼容CUDA及PyTorch、vLLM等主流推理框架。实测数据显示，在64K上下文场景下，单机Prefill吞吐达95,920 tok/s，按智谱API定价测算，满负载下单机月收入可达64.6万元；400K极限长序列场景TPM达2.6MTokens，吞吐平稳可控。摩尔线程表示，大模型产业竞争下半场本质上是推理效率与商业回报的较量，该方案旨在帮助企业在万亿参数与Agent时代降低总拥有成本。完整白皮书已在官网发布。