[{"data":1,"prerenderedAt":377},["ShallowReactive",2],{"content-doc-\u002Fzh\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery":3,"surround-\u002Fzh\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery":375},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"company":10,"date":11,"logo":12,"logoDark":13,"body":14,"_type":367,"_id":368,"_source":369,"_file":370,"_stem":371,"_extension":372,"coverImage":37,"plainText":373,"authorNames":374},"\u002Fzh\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery","usercase",false,"","基于 openUBMC 实现 PCIe Switch 多 GPU 自发现","随着大模型训练场景规模化落地，行业对单机高密度算力、低延迟互联、高稳定部署的需求持续提升。传统 CPU 直出架构受原生 PCIe\n通道物理限制，难以满足海量 GPU 挂载、高效算力互联的业务需求，存在带宽瓶颈、通信延迟高、组网灵活性不足、硬件设计成本高、故障易扩散等痛点，无法适配\nAI 训练场景对服务器高性能、高可靠、高灵活的严苛要求。","神州鲲泰","2026\u002F05\u002F29","\u002Fcategory\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery\u002Flogo.png","\u002Fcategory\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery\u002Flogo-dark.png",{"type":15,"children":16,"toc":360},"root",[17,25,30,40,45,50,57,296,301],{"type":18,"tag":19,"props":20,"children":22},"element","h3",{"id":21},"应用背景",[23],{"type":24,"value":21},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":24,"value":9},{"type":18,"tag":26,"props":31,"children":32},{},[33],{"type":18,"tag":34,"props":35,"children":39},"img",{"alt":36,"src":37,"title":38},"image","\u002Fcategory\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery\u002Fimg1.jpg","*该图仅用来解释switch的作用，并非实际拓扑",[],{"type":18,"tag":19,"props":41,"children":43},{"id":42},"解决方案",[44],{"type":24,"value":42},{"type":18,"tag":26,"props":46,"children":47},{},[48],{"type":24,"value":49},"神州鲲泰基于openUBMC 开源开放架构，依托其完善的硬件适配能力与标准化扩展框架，完成 PCIe Switch 板深度适配，实现 AI 服务器多\nGPU 智能自发现、统一管控与全链路硬件监控，打造适配高密度算力场景的 AI 服务器硬件管理方案。",{"type":18,"tag":26,"props":51,"children":52},{},[53],{"type":18,"tag":34,"props":54,"children":56},{"alt":36,"src":55},"\u002Fcategory\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery\u002Fimg2.jpg",[],{"type":18,"tag":58,"props":59,"children":60},"table",{},[61,80],{"type":18,"tag":62,"props":63,"children":64},"thead",{},[65],{"type":18,"tag":66,"props":67,"children":68},"tr",{},[69,75],{"type":18,"tag":70,"props":71,"children":72},"th",{},[73],{"type":24,"value":74},"关键特性",{"type":18,"tag":70,"props":76,"children":77},{},[78],{"type":24,"value":79},"规格描述",{"type":18,"tag":81,"props":82,"children":83},"tbody",{},[84,98,111,124,137,150,171,191,204,231,250,263,276],{"type":18,"tag":66,"props":85,"children":86},{},[87,93],{"type":18,"tag":88,"props":89,"children":90},"td",{},[91],{"type":24,"value":92},"产品形态",{"type":18,"tag":88,"props":94,"children":95},{},[96],{"type":24,"value":97},"机架式服务器",{"type":18,"tag":66,"props":99,"children":100},{},[101,106],{"type":18,"tag":88,"props":102,"children":103},{},[104],{"type":24,"value":105},"处理器",{"type":18,"tag":88,"props":107,"children":108},{},[109],{"type":24,"value":110},"2*鲲鹏920新型号 （48\u002F64\u002F核，最高主频2.9GHz）",{"type":18,"tag":66,"props":112,"children":113},{},[114,119],{"type":18,"tag":88,"props":115,"children":116},{},[117],{"type":24,"value":118},"AI加速卡",{"type":18,"tag":88,"props":120,"children":121},{},[122],{"type":24,"value":123},"最多支持10张全高全长双宽AI加速卡",{"type":18,"tag":66,"props":125,"children":126},{},[127,132],{"type":18,"tag":88,"props":128,"children":129},{},[130],{"type":24,"value":131},"内存",{"type":18,"tag":88,"props":133,"children":134},{},[135],{"type":24,"value":136},"最多支持32*DDR5 4800MT\u002Fs 及以上内存，支持ECC、热备、镜像等多种保护模式",{"type":18,"tag":66,"props":138,"children":139},{},[140,145],{"type":18,"tag":88,"props":141,"children":142},{},[143],{"type":24,"value":144},"存储控制器",{"type":18,"tag":88,"props":146,"children":147},{},[148],{"type":24,"value":149},"支持大容量缓存及超级电容。提供RAID级别迁移、磁盘漫游、自诊断、Web远程设置等功能，支持直通、RAID0、1、5、6、10、50、60等多种存储方案，支持RAID卡安全启动",{"type":18,"tag":66,"props":151,"children":152},{},[153,158],{"type":18,"tag":88,"props":154,"children":155},{},[156],{"type":24,"value":157},"本地存储",{"type":18,"tag":88,"props":159,"children":160},{},[161,163,169],{"type":24,"value":162},"前置：支持12",{"type":18,"tag":164,"props":165,"children":166},"em",{},[167],{"type":24,"value":168},"3.5\"盘位，其中8",{"type":24,"value":170},"SAS \u002F SATA+4*NVMe，支持热插拔",{"type":18,"tag":66,"props":172,"children":173},{},[174,179],{"type":18,"tag":88,"props":175,"children":176},{},[177],{"type":24,"value":178},"灵活插卡",{"type":18,"tag":88,"props":180,"children":181},{},[182,184,189],{"type":24,"value":183},"支持1",{"type":18,"tag":164,"props":185,"children":186},{},[187],{"type":24,"value":188},"OCP3.0标准网卡（4",{"type":24,"value":190},"GE电口 \u002F 2*25GE光口）",{"type":18,"tag":66,"props":192,"children":193},{},[194,199],{"type":18,"tag":88,"props":195,"children":196},{},[197],{"type":24,"value":198},"PCle槽位",{"type":18,"tag":88,"props":200,"children":201},{},[202],{"type":24,"value":203},"支持12*PCIe 5.0 x16扩展槽位",{"type":18,"tag":66,"props":205,"children":206},{},[207,212],{"type":18,"tag":88,"props":208,"children":209},{},[210],{"type":24,"value":211},"接口",{"type":18,"tag":88,"props":213,"children":214},{},[215,217,222,224,229],{"type":24,"value":216},"4",{"type":18,"tag":164,"props":218,"children":219},{},[220],{"type":24,"value":221},"USB 3.0接口，2",{"type":24,"value":223},"VGA接口，2",{"type":18,"tag":164,"props":225,"children":226},{},[227],{"type":24,"value":228},"UID按键&指示灯，1",{"type":24,"value":230},"串口、1*RJ45管理网口",{"type":18,"tag":66,"props":232,"children":233},{},[234,239],{"type":18,"tag":88,"props":235,"children":236},{},[237],{"type":24,"value":238},"电源模块",{"type":18,"tag":88,"props":240,"children":241},{},[242,243,248],{"type":24,"value":216},{"type":18,"tag":164,"props":244,"children":245},{},[246],{"type":24,"value":247},"热插拔CRPS 2600W电源，支持2+2冗余 \u002F 4",{"type":24,"value":249},"热插拔CRPS 2000W电源，支持3+1冗余",{"type":18,"tag":66,"props":251,"children":252},{},[253,258],{"type":18,"tag":88,"props":254,"children":255},{},[256],{"type":24,"value":257},"风扇模块",{"type":18,"tag":88,"props":259,"children":260},{},[261],{"type":24,"value":262},"配置12*颗热插拔风扇，支持单风扇失效状态下稳定工作",{"type":18,"tag":66,"props":264,"children":265},{},[266,271],{"type":18,"tag":88,"props":267,"children":268},{},[269],{"type":24,"value":270},"工作温度",{"type":18,"tag":88,"props":272,"children":273},{},[274],{"type":24,"value":275},"5℃~35℃",{"type":18,"tag":66,"props":277,"children":278},{},[279,284],{"type":18,"tag":88,"props":280,"children":281},{},[282],{"type":24,"value":283},"尺寸",{"type":18,"tag":88,"props":285,"children":286},{},[287,289,294],{"type":24,"value":288},"高176.8mm",{"type":18,"tag":164,"props":290,"children":291},{},[292],{"type":24,"value":293},"宽447.4mm",{"type":24,"value":295},"深871.4mm",{"type":18,"tag":19,"props":297,"children":299},{"id":298},"客户价值",[300],{"type":24,"value":298},{"type":18,"tag":302,"props":303,"children":304},"ol",{},[305,321,334,347],{"type":18,"tag":306,"props":307,"children":308},"li",{},[309,315,319],{"type":18,"tag":310,"props":311,"children":312},"strong",{},[313],{"type":24,"value":314},"突破硬件瓶颈，支撑高密度算力部署",{"type":18,"tag":316,"props":317,"children":318},"br",{},[],{"type":24,"value":320},"\n突破 CPU 原生 PCIe 通道物理上限，通过 PCIe Switch 实现通道高效拓展，满足单机海量高性能 GPU 挂载需求，完美支撑大模型训练的算力规模要求。",{"type":18,"tag":306,"props":322,"children":323},{},[324,329,332],{"type":18,"tag":310,"props":325,"children":326},{},[327],{"type":24,"value":328},"优化算力互联，大幅提升训练效率",{"type":18,"tag":316,"props":330,"children":331},{},[],{"type":24,"value":333},"\n实现 GPU 点对点低延迟直连通信，减少 CPU 中转带来的带宽损耗与数据延迟。相较传统 CPU 中转模式，带宽从 25GB\u002Fs 提升至 42GB\u002Fs，提升幅度约 70%，有效增强多卡算力协同效率，全面优化 AI 模型训练效果。",{"type":18,"tag":306,"props":335,"children":336},{},[337,342,345],{"type":18,"tag":310,"props":338,"children":339},{},[340],{"type":24,"value":341},"架构灵活可靠，适配多元业务需求",{"type":18,"tag":316,"props":343,"children":344},{},[],{"type":24,"value":346},"\n支持基于 PCIe Switch 的硬件拓扑灵活定制，可根据客户算力规模、业务场景、成本预算等差异，快速调整链路分配与总线规划，高效匹配多样化部署诉求。",{"type":18,"tag":306,"props":348,"children":349},{},[350,355,358],{"type":18,"tag":310,"props":351,"children":352},{},[353],{"type":24,"value":354},"简化硬件设计，降本增效",{"type":18,"tag":316,"props":356,"children":357},{},[],{"type":24,"value":359},"\n简化服务器主板 PCIe 高速走线设计，降低信号干扰与硬件故障风险，减少整机研发与生产成本，提升服务器规模化交付能力。",{"title":7,"searchDepth":361,"depth":361,"links":362},4,[363,365,366],{"id":21,"depth":364,"text":21},3,{"id":42,"depth":364,"text":42},{"id":298,"depth":364,"text":298},"markdown","content:zh:usercase:implementing-pice-switch-with-multi-gpu-self-discovery.md","content","zh\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery.md","zh\u002Fusercase\u002Fimplementing-pice-switch-with-multi-gpu-self-discovery","md","应用背景 随着大模型训练场景规模化落地，行业对单机高密度算力、低延迟互联、高稳定部署的需求持续提升。传统 CPU 直出架构受原生 PCIe\n通道物理限制，难以满足海量 GPU 挂载、高效算力互联的业务需求，存在带宽瓶颈、通信延迟高、组网灵活性不足、硬件设计成本高、故障易扩散等痛点，无法适配\nAI 训练场景对服务器高性能、高可靠、高灵活的严苛要求。  解决方案 神州鲲泰基于openUBMC 开源开放架构，依托其完善的硬件适配能力与标准化扩展框架，完成 PCIe Switch 板深度适配，实现 AI 服务器多\nGPU 智能自发现、统一管控与全链路硬件监控，打造适配高密度算力场景的 AI 服务器硬件管理方案。  关键特性 规格描述 产品形态 机架式服务器 处理器 2*鲲鹏920新型号 （48\u002F64\u002F核，最高主频2.9GHz） AI加速卡 最多支持10张全高全长双宽AI加速卡 内存 最多支持32*DDR5 4800MT\u002Fs 及以上内存，支持ECC、热备、镜像等多种保护模式 存储控制器 支持大容量缓存及超级电容。提供RAID级别迁移、磁盘漫游、自诊断、Web远程设置等功能，支持直通、RAID0、1、5、6、10、50、60等多种存储方案，支持RAID卡安全启动 本地存储 前置：支持12 3.5\"盘位，其中8 SAS \u002F SATA+4*NVMe，支持热插拔 灵活插卡 支持1 OCP3.0标准网卡（4 GE电口 \u002F 2*25GE光口） PCle槽位 支持12*PCIe 5.0 x16扩展槽位 接口 4 USB 3.0接口，2 VGA接口，2 UID按键&指示灯，1 串口、1*RJ45管理网口 电源模块 4 热插拔CRPS 2600W电源，支持2+2冗余 \u002F 4 热插拔CRPS 2000W电源，支持3+1冗余 风扇模块 配置12*颗热插拔风扇，支持单风扇失效状态下稳定工作 工作温度 5℃~35℃ 尺寸 高176.8mm 宽447.4mm 深871.4mm 客户价值 突破硬件瓶颈，支撑高密度算力部署  \n突破 CPU 原生 PCIe 通道物理上限，通过 PCIe Switch 实现通道高效拓展，满足单机海量高性能 GPU 挂载需求，完美支撑大模型训练的算力规模要求。 优化算力互联，大幅提升训练效率  \n实现 GPU 点对点低延迟直连通信，减少 CPU 中转带来的带宽损耗与数据延迟。相较传统 CPU 中转模式，带宽从 25GB\u002Fs 提升至 42GB\u002Fs，提升幅度约 70%，有效增强多卡算力协同效率，全面优化 AI 模型训练效果。 架构灵活可靠，适配多元业务需求  \n支持基于 PCIe Switch 的硬件拓扑灵活定制，可根据客户算力规模、业务场景、成本预算等差异，快速调整链路分配与总线规划，高效匹配多样化部署诉求。 简化硬件设计，降本增效  \n简化服务器主板 PCIe 高速走线设计，降低信号干扰与硬件故障风险，减少整机研发与生产成本，提升服务器规模化交付能力。",[],[376,376],null,1785154397653]