2026超节点技术白皮书

i 目 录 前言 ································· 1 超节点技术基础 ···························· 2 1 定义与核心特征 ····································· 2 1.1 超节点技术背景与演进必然性 ···························· 2 1.2 超节点架构演进 ·································· 7 1.3 超节点的技术定义与核心特征 ···························· 10 1.4 核心指标体系 ··································· 13 2 核心应用场景 ······································ 14 2.1 超大规模 AI 模型训练 ······························· 14 2.2 高性能计算（科学模拟、气象预测、药物研发） ···················· 15 2.3 关键行业 AI 基础设施（政务、金融、能源自主可控算力集群） ············· 16 2.4 混合负载场景（训练 + 推理协同、异构芯片集群） ··················· 17 超节点核心技术架构 ·························· 19 1 硬件架构设计 ······································ 19 1.1 超节点硬件架构 ·································· 19 1.2 硬件互联架构 ·································· 20 1.3 计算节点硬件设计 ································ 25 1.4 计算单元：算力芯片选型标准 ··························· 28 1.5 超节点互联拓扑 ································· 37 1.6 交换芯片技术 ·································· 45 ii 1.7 硬件集成方案 ·································· 46 2 Scale up 协议核心技术解析 ······························ 64 2.1 物理层及 SerDes 技术 ······························ 64 2.2 链路层 ····································· 74 2.3 事务层 ······································ 81 2.4 总结 ······································ 88 3 主要 Scale up 协议介绍 ································ 89 3.1 PCIe 与 CXL 协议介绍 ······························ 89 3.2 GLink 协议介绍 ································· 107 3.3 ETH-X 和 SUE/ESUN 协议介绍 ························· 120 3.4 OSIA 协议介绍 ································· 138 3.5 NVLink 协议介绍 ································ 161 3.6 UALink 协议介绍 ································ 171 3.7 参考文献 ···································· 209 软件栈与协同优化 ·························· 214 1 超节点软件栈 ····································· 214 2 软硬协同优化 ····································· 217 操作系统和超节点 ·························· 228 1 AI 工作负载场景与系统互联问题全景 ··························· 229 2 CPU-to-CPU 同构互联 ································ 232 3 CPU-to-GPU 异构互联 ································ 234 4 单节点多 GPU 互联 ·································· 238 5 超节点 GPU 互联 ··································· 240 6 资源池化 ······································· 251 7 互联语义演进：从 Read/Write 到统一 Load/Store 内存访问 ················ 253 8 总结与展望 ······································ 255 9 参考文献 ······································· 256 超节点技术选型与部署实践 ······················ 257 1 选型方法论 ······································ 257 iii 1.1 技术选型的意义 ································· 257 1.2 方法论整体框架 ································· 258 1.3 模型运行基础环境 ································ 258 1.4 业务指标（按训练/推理场景拆分） ························· 262 1.5 选型的实际案例 ································· 265 2 部署关键要求 ····································· 267 2.1 机房环境要求 ·································· 267 2.2 集群搭建流程 ·································· 276 3 运维与监控 ······································ 279 3.1 概述 ······································ 279 3.2 超节点运维核心范围 ············