[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-40":104},{"enabled":4,"title":5,"statusText":5,"buttonText":5,"campSlug":5,"stage":6,"classroomId":7,"hasClassroom":4},false,"",0,null,{"total":9,"list":10,"pageNum":44,"pageSize":45,"size":9,"startRow":44,"endRow":9,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":48,"navigateFirstPage":44,"navigateLastPage":44},3,[11,25,35],{"slug":12,"title":13,"logo":7,"liveUrl":14,"prize":15,"sponsor":16,"description":17,"organizer":18,"startTime":19,"endTime":20,"registrationDeadline":21,"status":22,"dashboardVisibility":23,"createdAt":24},"spring2026","2026春季启元人工智能大赛","https:\u002F\u002Fwww.bilibili.com\u002Fblackboard\u002Flive\u002Flive-activity-player.html?cid=1822703816&quality=0","200,000","麒麟软件有限公司","“2026春季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。\n各参赛选手需在规定的比赛周期内，依据赛事规则选定赛题并完成相应挑战，优胜者将共同角逐总额20万元的赛事奖金池，并且还将有机会获得实习和知名企业优质岗位内推资格。","启元实验室","2026-04-19 16:00:00","2026-07-13 16:00:00","2026-06-16 16:00:00","ongoing","private","2026-04-16 08:00:00",{"slug":26,"title":27,"logo":7,"liveUrl":14,"prize":28,"sponsor":7,"description":29,"organizer":18,"startTime":30,"endTime":31,"registrationDeadline":32,"status":33,"dashboardVisibility":23,"createdAt":34},"autumn2025","2025秋季启元人工智能大赛","150,000","“2025秋季启元人工智能大赛”是由启元实验室主办的面向全国人工智能领域开发者、科研人员及高校学生的权威赛事。大赛旨在促进与培养具备工程实践能力和创新能力的优秀人工智能领域开发者、推广国产软硬件技术与生态、助力国家人工智能产业高质量发展。","2025-11-08 08:00:00","2026-01-13 08:00:00","2025-12-30 08:00:00","ended","2025-11-03 08:00:00",{"slug":36,"title":37,"logo":7,"liveUrl":14,"prize":38,"sponsor":7,"description":39,"organizer":18,"startTime":40,"endTime":41,"registrationDeadline":42,"status":33,"dashboardVisibility":23,"createdAt":43},"summer2025","启元人工智能大赛(系统方向)","80,000","启元人工智能大赛由启元实验室主办，其中系统赛道基于启元实验室九源软件栈的相关开源项目，通过拉动开源社区力量，一方面促进国产智能计算生态的发展，另一方面也为人工智能系统方向的学生及从业者提供一个实践平台。","2025-07-08 08:00:00","2025-08-25 08:00:00","2025-07-22 08:00:00","2025-06-22 08:00:00",1,5,true,8,[44],{"total":45,"list":50,"pageNum":44,"pageSize":45,"size":45,"startRow":44,"endRow":45,"pages":44,"prePage":6,"nextPage":6,"isFirstPage":46,"isLastPage":46,"hasPreviousPage":4,"hasNextPage":4,"navigatePages":47,"navigatepageNums":103,"navigateFirstPage":44,"navigateLastPage":44},[51,62,73,84,94],{"id":52,"slug":53,"title":54,"logo":55,"publicLiveEmbedUrl":14,"description":56,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":16,"startTime":58,"endTime":59,"status":22,"dashboardVisibility":60,"createdAt":58,"updatedAt":61},"019eeeff-6b00-7d55-a629-efafded3e232","summer2026","2026年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2026-06\u002F019eef7f-7816-7828-9309-79c2c61dd568.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=pU5bMBWiYkrZ2T0DPHHg5hBG8fo%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。","...","2026-06-22 08:00:00","2026-10-19 08:00:00","public","2026-10-12 08:00:00",{"id":63,"slug":64,"title":65,"logo":66,"publicLiveEmbedUrl":7,"description":67,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":68,"startTime":69,"endTime":70,"status":33,"dashboardVisibility":23,"createdAt":71,"updatedAt":72},"019af3f3-e06a-7245-aa8d-37dc99a69c73","winter2025","2025年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-12\u002F019afc69-b7ca-7805-91fb-605429671c01.png?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jlKxcmMlK0FJ32tR4GVdVubbBQU%3D","由启元实验室发起的InfiniTensor大模型与人工智能系统训练营，面向开源社区、高校及产业界伙伴开放。活动以「普及AI全栈知识技术、培养工程实践与创新型人才、推广国产软硬件技术生态」为核心目标，助力国家人工智能产业高质量发展。 \n","新华三技术有限公司","2025-12-25 08:00:00","2026-03-27 08:00:00","2025-12-06 08:00:00","2026-10-22 20:24:30",{"id":74,"slug":36,"title":75,"logo":76,"publicLiveEmbedUrl":7,"description":77,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":78,"startTime":79,"endTime":80,"status":81,"dashboardVisibility":23,"createdAt":82,"updatedAt":83},"01980342-208c-7803-b3f3-19e1d31960ba","2025年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fuploads\u002F2025-07\u002F0198045e-28fe-78f0-95fe-23835ce58395.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J245EHlsDbmiAhfbqbOojujDz3M%3D","InfiniTensor大模型与人工智能系统训练营由启元实验室及清华大学 AI 系统方向团队发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营联合开源项目开发者、学术界大牛、工业界资深工程师参与授课与项目指导。课程采用「线上直播授课 + 在线答疑 + 项目实践辅导 」的方式进行，确保每位同学都能获得充分成长。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","北京积算科技有限公司","2025-07-14 08:00:00","2025-09-30 08:00:00","upcoming","2025-07-13 08:00:00","2026-10-22 20:24:58",{"id":85,"slug":86,"title":87,"logo":88,"publicLiveEmbedUrl":7,"description":89,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":90,"endTime":91,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":93},"01980cf9-a7d8-7796-ae86-0bbe05334add","winter2024","2024年冬季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-ce23-7624-a287-9f07dd0d85f6.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区、高校和产业界，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过视频、直播课、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。报名学员请阅读“导学”课程页的训练营介绍。","2025-01-05 08:00:00","2025-03-22 08:00:00","2025-07-15 08:00:00","2026-10-22 20:24:39",{"id":95,"slug":96,"title":97,"logo":98,"publicLiveEmbedUrl":7,"description":99,"introduction":57,"organizer":18,"coOrganizer":5,"sponsor":5,"startTime":100,"endTime":101,"status":81,"dashboardVisibility":23,"createdAt":92,"updatedAt":102},"01980c20-363a-7dc3-8f5a-58a68e08dcc4","summer2024","2024年夏季InfiniTensor大模型与人工智能系统训练营","https:\u002F\u002Ftesting-cdn.infinitensor.com\u002Fuploads\u002F2025-07\u002F01981269-fbfa-788a-afcd-17fa9be74ff9.jpg","InfiniTensor大模型与人工智能系统训练营由清华大学发起，面向开源社区和高校，普及人工智能各方面知识和技术，培养相关人才。训练营全程免费，通过课程、习题和项目实践，在人工智能飞速发展的背景下，为学员提供一窥深层技术的机会。","2024-07-20 08:00:00","2024-09-29 08:00:00","2026-10-22 20:24:49",[44],{"currentItem":105,"processedData":125},{"img":106,"keywords":107,"treeType":44,"description":108,"showStatus":44,"updateTime":109,"type":110,"title":111,"content":112,"nextArticle":113,"createTime":109,"sortOrder":117,"canonicalPath":118,"id":119,"contentType":120,"prevArticle":121},"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_fm.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TX4Q3eOa5B2qvukOh%2BxeoMTycMM%3D","InfiniTensor,冬季训练营,CUDA,异步并行,系统优化","聚焦 CUDA 异步并行与底层控制技巧，解析系统级性能优化思路与实战案例。","2026-10-22 20:38:11",2,"InfiniTensor 冬季训练营 CUDA 编程第六课《异步并行、底层控制与系统优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw01.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n9GJUuz%2B2gCcLDsyUFw2PeMavHY%3D)\r\n\r\n\r\n2026 年 1 月 27 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第六课《异步并行、底层控制与系统优化》开讲。\r\n\r\n本节课将深入讲解 CUDA 系统优化与高级技术，重点围绕延迟瓶颈的识别与优化展开。\r\n\r\n### 延迟瓶颈与优化基础\r\n#### 1.延迟瓶颈（Latency-bounded）概念:\r\n- 计算与访存利用率低，线程因等待长延迟操作而阻塞\r\n#### 2.性能瓶颈识别方法:\r\n- 使用 NCU 分析工具识别瓶颈指标（如 Stall Long Scoreboard）\r\n#### 3.硬件单元独立性:\r\n- 计算单元与访存单元属于不同的硬件执行单元，在满足依赖与调度条件时可实现一定程度的并行重叠。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw02.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YFXIym595SpkYNbUZY0WZl6UoWw%3D)\r\n\r\n### 存算重叠技术\r\n#### 1.核心概念:\r\n- 访存与计算的时间重叠，实现延迟隐藏（Latency Hiding）\r\n#### 2.实现方法:\r\n- 双缓冲（Double Buffering）：设置两个共享内存缓冲区\r\n- 线程束特化（Warp Specialization）：不同 warps 承担不同角色（计算\u002F访存）\r\n#### 3.硬件基础:\r\n- SM 内多流水线并行执行能力\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw03.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RdH1KQVbS4JzjGOyMyr7HT5gHtU%3D)\r\n\r\n### 异步执行与流水线\r\n#### 1.异步编程模型:\r\n- 核函数发射的异步特性\r\n- CUDA Pipelining API\r\n- memcpy_async 异步数据拷贝\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw04.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yUuvhheP%2Fi5gw9qBv4neMnitNB0%3D)\r\n\r\n#### 2.生产者-消费者模型:\r\n- Pipelining 工作流程：Creates→Acquire→Submit→Commit→Wait→Compute→Release\r\n#### 3.性能对比:\r\n- 异步加载相比同步加载有显著性能提升\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw05.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SjtBzqLak%2Bp%2FTcahMIws%2Fkfgue8%3D)\r\n\r\n### 多流并行技术\r\n#### 1.流（Stream）概念:\r\n- CUDA 中的执行队列（FIFO）\r\n#### 2.流类型​\r\n- 默认流（Default Stream）：在 Legacy 模式下与所有其他流隐式同步\r\n- 显示流（Named Streams）：可创建的非阻塞流\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw06.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=gLeVw9ykHmTqvOxivaFeEPPWsyc%3D)\r\n\r\n#### 3.多流并行\r\n- 使用 cudaStreamCreateWithFlags 创建非阻塞流\r\n- 通过 NVIDIA Nsight Systems 分析时间线\r\n#### 4.异步内存操作​\r\n- cudaMallocAsync\u002FcudaMemcpyAsync 异步分配和拷贝\r\n- 固定内存（Pinned Memory）与零拷贝技术\r\n\r\n### 统一内存管理\r\n#### 1.统一内存（Unified Memory）:\r\n- CPU\u002FGPU 共享内存空间\r\n#### 2.管理策略:\r\n- cudaMallocManaged 自动数据迁移\r\n- 基于 Page Fault 的换页机制\r\n- 优化 API：cudaMemAdvise 启发式策略、cudaMemPrefetchAsync 异步预取\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw07.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8YiBsGJBdV71x3%2Fufx81Izsdm%2F4%3D)\r\n\r\n### 计算图优化\r\n#### 1.CUDA Graph 概念:\r\n- 将操作序列组织为有向无环图\r\n#### 2.核心优势​\r\n- 减少核函数启动开销\r\n- 增强任务并行性与依赖性识别\r\n- 避免重复工作\r\n#### 3.工作流程​\r\n- 流捕获→图创建→实例化→启动执行\r\n#### 4.性能表现\r\n- 批量提交操作，消除执行间隙（Bubble）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw08.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0AeO87WZG3idJn7ukRj2Y4NtAIU%3D)\r\n\r\n### 多 GPU 协同计算\r\n#### 1.硬件互联结构:\r\n- PCIe 树状连接 vs NVLink 网状连接\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw09.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2FpFXiCpKKcEtIaAJSD2gUKqlSs%3D)\r\n\r\n#### 2.协作方式​\r\n- 点对点传输（P2P）与 RDMA 直接内存访问\r\n- 统一内存跨设备访问\r\n- NCCL 集合通信库\r\n#### 3.通信模式​\r\n- 点对点通信（Send\u002FReceive）\r\n- 集合通信（AllReduce等）\r\n- 多进程\u002F多线程环境下的 GPU 管理\r\n\r\n### 优化指导原则\r\n#### 1.性能定律​\r\n- 阿姆达尔定律：串行部分限制加速上限\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw10.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MOKlcTTzmtmVsyE2AxxtUmFBevg%3D)\r\n\r\n- 古斯塔夫森定律：扩大问题规模提升实际加速\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw11.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=gvDd2W7za32DCDyraaxBXpzvZfc%3D)\r\n\r\n#### 2.优化哲学​\r\n- 优先解决性能瓶颈最大的环节，避免盲目优化\r\n\r\n### 总结\r\n本节课深入探讨了 CUDA 系统级优化与高级技术，重点围绕延迟瓶颈的识别与优化展开，从单一 GPU 优化扩展到多 GPU 协同计算，下节课将详细介绍更加快速性能提升的方法，和高阶的优化方法。\r\n\r\n\r\n\r\n\r\n![2025冬季训练营课程表](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2F7TZ1TuYp4PCgB34LK2cfXydhgo%3D)\r\n\r\n### 💡 学习贴士\r\n更多细节及完整课程内容，请观看直播或查看课程回放。\r\n\r\n链接：https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\r\n\r\n### 📚 训练营报名持续进行中\r\n扫码报名与加入群聊，或点击 **「阅读原文」** 直达官网报名\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6ZWvZ8NpGN%2BxarpfEM4E7cFRad4%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-39","study-39","InfiniTensor 冬季训练营 Triton&九齿第五课《天数智芯 Triton 实战》",69,"\u002Fstudy-material\u002Fstudy-40","study-40","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-41","study-41","InfiniTensor 冬季训练营 AI 编译器第六课《Stable Diffusion 在 InfiniTensor 推理框架中的适配与工程实践》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw01.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=n9GJUuz%2B2gCcLDsyUFw2PeMavHY%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2026 年 1 月 27 日，2025 冬季 InfiniTensor 大模型与人工智能系统训练营 CUDA 编程系列第六课《异步并行、底层控制与系统优化》开讲。\u003C\u002Fp>\n\u003Cp>本节课将深入讲解 CUDA 系统优化与高级技术，重点围绕延迟瓶颈的识别与优化展开。\u003C\u002Fp>\n\u003Ch3 id=\"延迟瓶颈与优化基础\">\u003Ca class=\"anchor\" href=\"#延迟瓶颈与优化基础\" aria-label=\"跳转到标题：延迟瓶颈与优化基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>延迟瓶颈与优化基础\u003C\u002Fh3>\u003Ch4 id=\"1-延迟瓶颈-latency-bounded-概念\">\u003Ca class=\"anchor\" href=\"#1-延迟瓶颈-latency-bounded-概念\" aria-label=\"跳转到标题：1.延迟瓶颈（Latency-bounded）概念:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.延迟瓶颈（Latency-bounded）概念:\u003C\u002Fh4>\u003Cul>\n\u003Cli>计算与访存利用率低，线程因等待长延迟操作而阻塞\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-性能瓶颈识别方法\">\u003Ca class=\"anchor\" href=\"#2-性能瓶颈识别方法\" aria-label=\"跳转到标题：2.性能瓶颈识别方法:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.性能瓶颈识别方法:\u003C\u002Fh4>\u003Cul>\n\u003Cli>使用 NCU 分析工具识别瓶颈指标（如 Stall Long Scoreboard）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-硬件单元独立性\">\u003Ca class=\"anchor\" href=\"#3-硬件单元独立性\" aria-label=\"跳转到标题：3.硬件单元独立性:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.硬件单元独立性:\u003C\u002Fh4>\u003Cul>\n\u003Cli>计算单元与访存单元属于不同的硬件执行单元，在满足依赖与调度条件时可实现一定程度的并行重叠。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw02.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=YFXIym595SpkYNbUZY0WZl6UoWw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"存算重叠技术\">\u003Ca class=\"anchor\" href=\"#存算重叠技术\" aria-label=\"跳转到标题：存算重叠技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>存算重叠技术\u003C\u002Fh3>\u003Ch4 id=\"1-核心概念\">\u003Ca class=\"anchor\" href=\"#1-核心概念\" aria-label=\"跳转到标题：1.核心概念:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.核心概念:\u003C\u002Fh4>\u003Cul>\n\u003Cli>访存与计算的时间重叠，实现延迟隐藏（Latency Hiding）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-实现方法\">\u003Ca class=\"anchor\" href=\"#2-实现方法\" aria-label=\"跳转到标题：2.实现方法:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.实现方法:\u003C\u002Fh4>\u003Cul>\n\u003Cli>双缓冲（Double Buffering）：设置两个共享内存缓冲区\u003C\u002Fli>\n\u003Cli>线程束特化（Warp Specialization）：不同 warps 承担不同角色（计算\u002F访存）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-硬件基础\">\u003Ca class=\"anchor\" href=\"#3-硬件基础\" aria-label=\"跳转到标题：3.硬件基础:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.硬件基础:\u003C\u002Fh4>\u003Cul>\n\u003Cli>SM 内多流水线并行执行能力\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw03.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RdH1KQVbS4JzjGOyMyr7HT5gHtU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"异步执行与流水线\">\u003Ca class=\"anchor\" href=\"#异步执行与流水线\" aria-label=\"跳转到标题：异步执行与流水线\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>异步执行与流水线\u003C\u002Fh3>\u003Ch4 id=\"1-异步编程模型\">\u003Ca class=\"anchor\" href=\"#1-异步编程模型\" aria-label=\"跳转到标题：1.异步编程模型:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.异步编程模型:\u003C\u002Fh4>\u003Cul>\n\u003Cli>核函数发射的异步特性\u003C\u002Fli>\n\u003Cli>CUDA Pipelining API\u003C\u002Fli>\n\u003Cli>memcpy_async 异步数据拷贝\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw04.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=yUuvhheP%2Fi5gw9qBv4neMnitNB0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-生产者-消费者模型\">\u003Ca class=\"anchor\" href=\"#2-生产者-消费者模型\" aria-label=\"跳转到标题：2.生产者-消费者模型:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.生产者-消费者模型:\u003C\u002Fh4>\u003Cul>\n\u003Cli>Pipelining 工作流程：Creates→Acquire→Submit→Commit→Wait→Compute→Release\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-性能对比\">\u003Ca class=\"anchor\" href=\"#3-性能对比\" aria-label=\"跳转到标题：3.性能对比:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.性能对比:\u003C\u002Fh4>\u003Cul>\n\u003Cli>异步加载相比同步加载有显著性能提升\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw05.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SjtBzqLak%2Bp%2FTcahMIws%2Fkfgue8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"多流并行技术\">\u003Ca class=\"anchor\" href=\"#多流并行技术\" aria-label=\"跳转到标题：多流并行技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>多流并行技术\u003C\u002Fh3>\u003Ch4 id=\"1-流-stream-概念\">\u003Ca class=\"anchor\" href=\"#1-流-stream-概念\" aria-label=\"跳转到标题：1.流（Stream）概念:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.流（Stream）概念:\u003C\u002Fh4>\u003Cul>\n\u003Cli>CUDA 中的执行队列（FIFO）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-流类型\">\u003Ca class=\"anchor\" href=\"#2-流类型\" aria-label=\"跳转到标题：2.流类型​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.流类型​\u003C\u002Fh4>\u003Cul>\n\u003Cli>默认流（Default Stream）：在 Legacy 模式下与所有其他流隐式同步\u003C\u002Fli>\n\u003Cli>显示流（Named Streams）：可创建的非阻塞流\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw06.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=gLeVw9ykHmTqvOxivaFeEPPWsyc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"3-多流并行\">\u003Ca class=\"anchor\" href=\"#3-多流并行\" aria-label=\"跳转到标题：3.多流并行\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.多流并行\u003C\u002Fh4>\u003Cul>\n\u003Cli>使用 cudaStreamCreateWithFlags 创建非阻塞流\u003C\u002Fli>\n\u003Cli>通过 NVIDIA Nsight Systems 分析时间线\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"4-异步内存操作\">\u003Ca class=\"anchor\" href=\"#4-异步内存操作\" aria-label=\"跳转到标题：4.异步内存操作​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4.异步内存操作​\u003C\u002Fh4>\u003Cul>\n\u003Cli>cudaMallocAsync\u002FcudaMemcpyAsync 异步分配和拷贝\u003C\u002Fli>\n\u003Cli>固定内存（Pinned Memory）与零拷贝技术\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"统一内存管理\">\u003Ca class=\"anchor\" href=\"#统一内存管理\" aria-label=\"跳转到标题：统一内存管理\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>统一内存管理\u003C\u002Fh3>\u003Ch4 id=\"1-统一内存-unified-memory\">\u003Ca class=\"anchor\" href=\"#1-统一内存-unified-memory\" aria-label=\"跳转到标题：1.统一内存（Unified Memory）:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.统一内存（Unified Memory）:\u003C\u002Fh4>\u003Cul>\n\u003Cli>CPU\u002FGPU 共享内存空间\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-管理策略\">\u003Ca class=\"anchor\" href=\"#2-管理策略\" aria-label=\"跳转到标题：2.管理策略:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.管理策略:\u003C\u002Fh4>\u003Cul>\n\u003Cli>cudaMallocManaged 自动数据迁移\u003C\u002Fli>\n\u003Cli>基于 Page Fault 的换页机制\u003C\u002Fli>\n\u003Cli>优化 API：cudaMemAdvise 启发式策略、cudaMemPrefetchAsync 异步预取\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw07.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8YiBsGJBdV71x3%2Fufx81Izsdm%2F4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"计算图优化\">\u003Ca class=\"anchor\" href=\"#计算图优化\" aria-label=\"跳转到标题：计算图优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算图优化\u003C\u002Fh3>\u003Ch4 id=\"1-cuda-graph-概念\">\u003Ca class=\"anchor\" href=\"#1-cuda-graph-概念\" aria-label=\"跳转到标题：1.CUDA Graph 概念:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.CUDA Graph 概念:\u003C\u002Fh4>\u003Cul>\n\u003Cli>将操作序列组织为有向无环图\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-核心优势\">\u003Ca class=\"anchor\" href=\"#2-核心优势\" aria-label=\"跳转到标题：2.核心优势​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.核心优势​\u003C\u002Fh4>\u003Cul>\n\u003Cli>减少核函数启动开销\u003C\u002Fli>\n\u003Cli>增强任务并行性与依赖性识别\u003C\u002Fli>\n\u003Cli>避免重复工作\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-工作流程\">\u003Ca class=\"anchor\" href=\"#3-工作流程\" aria-label=\"跳转到标题：3.工作流程​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.工作流程​\u003C\u002Fh4>\u003Cul>\n\u003Cli>流捕获→图创建→实例化→启动执行\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"4-性能表现\">\u003Ca class=\"anchor\" href=\"#4-性能表现\" aria-label=\"跳转到标题：4.性能表现\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4.性能表现\u003C\u002Fh4>\u003Cul>\n\u003Cli>批量提交操作，消除执行间隙（Bubble）\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw08.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0AeO87WZG3idJn7ukRj2Y4NtAIU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"多-gpu-协同计算\">\u003Ca class=\"anchor\" href=\"#多-gpu-协同计算\" aria-label=\"跳转到标题：多 GPU 协同计算\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>多 GPU 协同计算\u003C\u002Fh3>\u003Ch4 id=\"1-硬件互联结构\">\u003Ca class=\"anchor\" href=\"#1-硬件互联结构\" aria-label=\"跳转到标题：1.硬件互联结构:\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.硬件互联结构:\u003C\u002Fh4>\u003Cul>\n\u003Cli>PCIe 树状连接 vs NVLink 网状连接\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw09.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2FpFXiCpKKcEtIaAJSD2gUKqlSs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-协作方式\">\u003Ca class=\"anchor\" href=\"#2-协作方式\" aria-label=\"跳转到标题：2.协作方式​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.协作方式​\u003C\u002Fh4>\u003Cul>\n\u003Cli>点对点传输（P2P）与 RDMA 直接内存访问\u003C\u002Fli>\n\u003Cli>统一内存跨设备访问\u003C\u002Fli>\n\u003Cli>NCCL 集合通信库\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-通信模式\">\u003Ca class=\"anchor\" href=\"#3-通信模式\" aria-label=\"跳转到标题：3.通信模式​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3.通信模式​\u003C\u002Fh4>\u003Cul>\n\u003Cli>点对点通信（Send\u002FReceive）\u003C\u002Fli>\n\u003Cli>集合通信（AllReduce等）\u003C\u002Fli>\n\u003Cli>多进程\u002F多线程环境下的 GPU 管理\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"优化指导原则\">\u003Ca class=\"anchor\" href=\"#优化指导原则\" aria-label=\"跳转到标题：优化指导原则\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>优化指导原则\u003C\u002Fh3>\u003Ch4 id=\"1-性能定律\">\u003Ca class=\"anchor\" href=\"#1-性能定律\" aria-label=\"跳转到标题：1.性能定律​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1.性能定律​\u003C\u002Fh4>\u003Cul>\n\u003Cli>阿姆达尔定律：串行部分限制加速上限\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw10.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MOKlcTTzmtmVsyE2AxxtUmFBevg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>古斯塔夫森定律：扩大问题规模提升实际加速\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_40_zw11.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=gvDd2W7za32DCDyraaxBXpzvZfc%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"2-优化哲学\">\u003Ca class=\"anchor\" href=\"#2-优化哲学\" aria-label=\"跳转到标题：2.优化哲学​\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2.优化哲学​\u003C\u002Fh4>\u003Cul>\n\u003Cli>优先解决性能瓶颈最大的环节，避免盲目优化\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"总结\">\u003Ca class=\"anchor\" href=\"#总结\" aria-label=\"跳转到标题：总结\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>总结\u003C\u002Fh3>\u003Cp>本节课深入探讨了 CUDA 系统级优化与高级技术，重点围绕延迟瓶颈的识别与优化展开，从单一 GPU 优化扩展到多 GPU 协同计算，下节课将详细介绍更加快速性能提升的方法，和高阶的优化方法。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_kcb.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2F7TZ1TuYp4PCgB34LK2cfXydhgo%3D\" alt=\"2025冬季训练营课程表\">\u003C\u002Fp>\n\u003Ch3 id=\"学习贴士\">\u003Ca class=\"anchor\" href=\"#学习贴士\" aria-label=\"跳转到标题：💡 学习贴士\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>💡 学习贴士\u003C\u002Fh3>\u003Cp>更多细节及完整课程内容，请观看直播或查看课程回放。\u003C\u002Fp>\n\u003Cp>链接：\u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\">https:\u002F\u002Fwww.infinitensor.com\u002Fcamp\u002Fwinter2025\u003C\u002Fa>\u003C\u002Fp>\n\u003Ch3 id=\"训练营报名持续进行中\">\u003Ca class=\"anchor\" href=\"#训练营报名持续进行中\" aria-label=\"跳转到标题：📚 训练营报名持续进行中\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📚 训练营报名持续进行中\u003C\u002Fh3>\u003Cp>扫码报名与加入群聊，或点击 \u003Cstrong>「阅读原文」\u003C\u002Fstrong> 直达官网报名\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_ewm.jpg?Expires=1792752558&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6ZWvZ8NpGN%2BxarpfEM4E7cFRad4%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,134,137,140,142,145,148,151,153,156,159,162,164,167,170,173,176,178,181,184,186,189,192,195,198,201,204,207,210,212,215,218,220,223],{"id":129,"text":129,"level":9},"延迟瓶颈与优化基础",{"id":131,"text":132,"level":133},"1-延迟瓶颈-latency-bounded-概念","1.延迟瓶颈（Latency-bounded）概念:",4,{"id":135,"text":136,"level":133},"2-性能瓶颈识别方法","2.性能瓶颈识别方法:",{"id":138,"text":139,"level":133},"3-硬件单元独立性","3.硬件单元独立性:",{"id":141,"text":141,"level":9},"存算重叠技术",{"id":143,"text":144,"level":133},"1-核心概念","1.核心概念:",{"id":146,"text":147,"level":133},"2-实现方法","2.实现方法:",{"id":149,"text":150,"level":133},"3-硬件基础","3.硬件基础:",{"id":152,"text":152,"level":9},"异步执行与流水线",{"id":154,"text":155,"level":133},"1-异步编程模型","1.异步编程模型:",{"id":157,"text":158,"level":133},"2-生产者-消费者模型","2.生产者-消费者模型:",{"id":160,"text":161,"level":133},"3-性能对比","3.性能对比:",{"id":163,"text":163,"level":9},"多流并行技术",{"id":165,"text":166,"level":133},"1-流-stream-概念","1.流（Stream）概念:",{"id":168,"text":169,"level":133},"2-流类型","2.流类型​",{"id":171,"text":172,"level":133},"3-多流并行","3.多流并行",{"id":174,"text":175,"level":133},"4-异步内存操作","4.异步内存操作​",{"id":177,"text":177,"level":9},"统一内存管理",{"id":179,"text":180,"level":133},"1-统一内存-unified-memory","1.统一内存（Unified Memory）:",{"id":182,"text":183,"level":133},"2-管理策略","2.管理策略:",{"id":185,"text":185,"level":9},"计算图优化",{"id":187,"text":188,"level":133},"1-cuda-graph-概念","1.CUDA Graph 概念:",{"id":190,"text":191,"level":133},"2-核心优势","2.核心优势​",{"id":193,"text":194,"level":133},"3-工作流程","3.工作流程​",{"id":196,"text":197,"level":133},"4-性能表现","4.性能表现",{"id":199,"text":200,"level":9},"多-gpu-协同计算","多 GPU 协同计算",{"id":202,"text":203,"level":133},"1-硬件互联结构","1.硬件互联结构:",{"id":205,"text":206,"level":133},"2-协作方式","2.协作方式​",{"id":208,"text":209,"level":133},"3-通信模式","3.通信模式​",{"id":211,"text":211,"level":9},"优化指导原则",{"id":213,"text":214,"level":133},"1-性能定律","1.性能定律​",{"id":216,"text":217,"level":133},"2-优化哲学","2.优化哲学​",{"id":219,"text":219,"level":9},"总结",{"id":221,"text":222,"level":9},"学习贴士","💡 学习贴士",{"id":224,"text":225,"level":9},"训练营报名持续进行中","📚 训练营报名持续进行中"]