[{"data":1,"prerenderedAt":132},["ShallowReactive",2],{"content-doc-\u002Fzh\u002Fnews\u002F20261026":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"author":11,"tags":12,"body":15,"_type":126,"_id":127,"_source":128,"_file":129,"_stem":130,"_extension":131},"\u002Fzh\u002Fnews\u002F20261026","news",false,"","聚焦低精度计算创新，共探大模型高效落地路径——GCC-HiFloat社区在成都CNCC 2026举行低精度计算论坛","2026年10月22日，GCC-HiFloat社区在CNCC 2026举办低精度计算论坛。来自高校、科研院所及企业的专家围绕低精度数值机理、数据格式、模型量化与软硬件协同优化分享研究成果和实践经验，探讨大模型训练与推理的效率提升路径。","2026\u002F10\u002F26","全球计算联盟GCC",[13,14],"CNCC","论坛",{"type":16,"children":17,"toc":123},"root",[18,25,34,39,46,51,58,63,70,75,82,87,94,99,106,111,118],{"type":19,"tag":20,"props":21,"children":22},"element","p",{},[23],{"type":24,"value":9},"text",{"type":19,"tag":20,"props":26,"children":27},{},[28],{"type":19,"tag":29,"props":30,"children":33},"img",{"alt":31,"src":32},"0","\u002Fnews\u002F20261026\u002F0.jpg",[],{"type":19,"tag":20,"props":35,"children":36},{},[37],{"type":24,"value":38},"华南理工大学陆璐教授分享《低精度量化算子开发与性能优化》，以低精度矩阵乘和FlashAttention为例，剖析缩放因子、访存及同步开销带来的瓶颈。他指出，降低位宽后，非矩阵计算开销更加突出，应通过性能建模、消融分析与流水优化定位瓶颈，并推动硬件资源配比协同改进。",{"type":19,"tag":20,"props":40,"children":41},{},[42],{"type":19,"tag":29,"props":43,"children":45},{"alt":31,"src":44},"\u002Fnews\u002F20261026\u002F1.jpg",[],{"type":19,"tag":20,"props":47,"children":48},{},[49],{"type":24,"value":50},"复旦大学尚笠教授作《大模型数值精度：机理与瓶颈》报告，从谱各向异性出发，分析量化对长尾信息和任务敏感方向的影响，介绍Metis与Averis等方法。他强调，整体量化误差不足以衡量模型损伤，需关注系统性偏差及误差方向，保护弱谱信息，探索无损低精度学习。",{"type":19,"tag":20,"props":52,"children":53},{},[54],{"type":19,"tag":29,"props":55,"children":57},{"alt":31,"src":56},"\u002Fnews\u002F20261026\u002F2.jpg",[],{"type":19,"tag":20,"props":59,"children":60},{},[61],{"type":24,"value":62},"中国科学院自动化研究所王培松研究院作《大模型量化压缩与高效部署》报告，介绍逐层旋转校准、分块FP4量化、多模态混合精度及自适应卸载等成果。他指出，压缩方法应适配数据分布与硬件特性，在保持模型精度的同时进一步关注推理行为，通过算法、算子与硬件协同提升部署效率。",{"type":19,"tag":20,"props":64,"children":65},{},[66],{"type":19,"tag":29,"props":67,"children":69},{"alt":31,"src":68},"\u002Fnews\u002F20261026\u002F3.jpg",[],{"type":19,"tag":20,"props":71,"children":72},{},[73],{"type":24,"value":74},"南京大学鲁金铭研究员就《面向AI计算的低精度数据格式演进与软硬件协同优化》做出分享，梳理从通用浮点到层次化缩放格式的演进，分享分段整数、对数块浮点及查表计算等探索。他强调，应统筹编码空间、缩放粒度与数据分布，联合优化格式、算法、编译和架构，将位宽优势转化为系统能效。",{"type":19,"tag":20,"props":76,"children":77},{},[78],{"type":19,"tag":29,"props":79,"children":81},{"alt":31,"src":80},"\u002Fnews\u002F20261026\u002F4.jpg",[],{"type":19,"tag":20,"props":83,"children":84},{},[85],{"type":24,"value":86},"上海交通大学冷静文教授作《极低比特大模型推理：从数值表示到高效执行》报告，介绍自适应数值格式、元数据补偿与真实硬件数值仿真等研究。他指出，解码与预填充需分别应对访存和计算瓶颈，量化评估还应对齐硬件累加、舍入及归约行为，以格式与执行协同支撑极低比特推理。",{"type":19,"tag":20,"props":88,"children":89},{},[90],{"type":19,"tag":29,"props":91,"children":93},{"alt":31,"src":92},"\u002Fnews\u002F20261026\u002F5.jpg",[],{"type":19,"tag":20,"props":95,"children":96},{},[97],{"type":24,"value":98},"香港理工大学秦浩桐就《迈向极致低精度的基础模型：从数值、张量到时序》展开分享，从数值量化、张量压缩与时序复用三个维度介绍研究进展，分享2比特量化、分阶段视频量化及特征复用等方法。他提出，应保留关键信息、压缩冗余表示并减少重复计算，兼顾模型能力与实际加速。",{"type":19,"tag":20,"props":100,"children":101},{},[102],{"type":19,"tag":29,"props":103,"children":105},{"alt":31,"src":104},"\u002Fnews\u002F20261026\u002F6.jpg",[],{"type":19,"tag":20,"props":107,"children":108},{},[109],{"type":24,"value":110},"华为技术有限公司李思坦作《HiFloat数据格式与实践案例》报告，介绍HiF8锥形精度、HiF4三级共享缩放及训练推理实践。她指出，合理缩放是发挥格式优势的关键，结合算子融合与连续低精度数据流，可减少格式转换和数据搬运开销，推动优化从单算子加速走向整网性能提升。",{"type":19,"tag":20,"props":112,"children":113},{},[114],{"type":19,"tag":29,"props":115,"children":117},{"alt":31,"src":116},"\u002Fnews\u002F20261026\u002F7.jpg",[],{"type":19,"tag":20,"props":119,"children":120},{},[121],{"type":24,"value":122},"本次论坛集中呈现了低精度计算从数值理论、格式设计到工程部署的多层次探索，为产学研交流提供了平台。各项报告表明，低精度计算的持续发展需要兼顾信息保真与执行效率，通过软硬件协同，将更低位宽转化为可验证的性能与能效收益。",{"title":7,"searchDepth":124,"depth":124,"links":125},2,[],"markdown","app:zh:news:20261026.md","app","zh\u002Fnews\u002F20261026.md","zh\u002Fnews\u002F20261026","md",1795070546055]