您的位置：首页 > 新闻中心 > 社会 > 正文

焦糖

摩尔线程发布技术白皮书：破解长上下文推理成本瓶颈_我的网站

勇者行动

一 | (ECNS)-- The third phase of an eco-agrotourism project under the “Yunnan Hand in Hand” initiative was launched in Bali, Indonesia, on Wednesday, following the signing of a new cooperation agreement between Yunnan and Bali. The project will install 27 solar-powered street lights and establish 10 hectares of demonstration zones for organic rice cultivation in Tabanan Regency. Luh Ayu Aryani, assistant for economy and development at the Bali Provincial Regional Secretariat, noted that the project aligns with the province’s sustainable development needs and creates new opportunities for farmers and small businesses by increasing the value of agricultural products. The “Yunnan Hand in Hand” initiative is a public-interest program targeting countries in South Asia, Southeast Asia and the Indian Ocean region. Since its launch in Bali in 2024, the program has benefited more than 1,000 local villagers. (By Helen Mo, intern Lin Qiaochu) 。

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。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。