[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-80":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=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fLKIdTmv5ztTF5lNyW4vFO%2FwiU4%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=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=IfcKyBkOe6Ug2cb8ODO9I1IUNkA%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=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ev4DWdZSXJ0t1T4u2JL5qybAazs%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_80_fm.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PuyaBRpLRhmOFlef0LgrCtamz1E%3D","CUDA,SIMT,向量规约,Shared Memory,Warp Shuffle","讲解 CPU 与 GPU 硬件差异、多层存储与 SM 架构，完整介绍 Grid\u002FBlock\u002FThread 编程模型，实战向量加法与分层树状规约，补充 Warp Shuffle、CUB 库加速手段。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 CUDA 编程 4：《CUDA 基础入门：向量规约》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fkVCF%2Bq28sfHht31UdWLnD7Qk8I%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 27 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第四课《CUDA 基础入门：向量规约》开讲。\r\n前三课依次讲解 CPU \u002F GPU 架构差异、SIMT 基础概念、 Roofline 性能模型、GPU 多层内存架构、Shared Memory Bank Conflict，以及原子操作、Warp \u002F 树状规约、Cooperative Groups 等多级并行规约优化方案。\u003C\u002Fspan>\r\n\r\n本节课从 CPU \u002F GPU 硬件差异切入，讲解 CUDA 完整编程模型、向量基础算子与并行向量规约实现，打通 CUDA 底层硬件到并行代码落地全链路。\r\n\r\n\r\n\r\n## CPU 与 GPU 硬件核心差异\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MhdJ9diLJonAqMIrSfnvCRq1PXE%3D)\r\n\r\n- CPU：大量缓存、控制单元，计算单元少，适配复杂分支逻辑；\r\n- GPU 海量计算单元，缓存占比极低，面向大批量连续并行数据运算。\r\n- 关键区别：CPU 依赖 L3 缓存复用重复数据，GPU 依靠 Warp 切换掩盖访存延迟，不采用类似 CPU 的大容量共享 L3 Cache 设计。\r\n- Warp 由 32 个线程组成，GPU 为每个 Warp 独立分配寄存器；GPU 一个 SM 可以同时驻留多个 Warp，Warp Scheduler 会在每个周期选择满足执行条件的 Warp 发射指令。当某个 Warp 因访存等原因等待时，调度器可以切换执行其他 ready Warp，从而隐藏访存延迟。\r\n\r\n## GPU 完整存储层级与性能差距\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XSz4ISn6QAnjQViUKJcQ22n0kiU%3D)\r\n\r\n### GPU 存储层级（速度由快至慢）\r\n1. **寄存器（RMEM）**：线程私有，常见单线程上限 255 个，变量直接映射寄存器；\r\n2. **Shared Memory（SMEM）**：单 Block 线程共享，可编程，存在 Bug Conflict 性能隐患；\r\n3. 显存（GMEM）：全局大容量存储。在大模型推理尤其是 Decode 阶段，显存带宽通常是影响性能的重要指标之一。\r\n\r\n补充：L2 Cache 硬件自动缓存，不可手动控制，极限优化场景少量利用。\r\n\r\n### 异步访存硬件革新\r\nGPU 支持显存 ↔ Shared Memory 异步传输，访存与计算可并行执行，大幅掩盖内存延迟；新一代 TMA 硬件进一步节省寄存器开销。\r\n\r\n## GPU 的计算单元\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ydvma2%2B5IR%2Bn4IRLeGQhEkfGWMk%3D)\r\n\r\n#### 1. SM 基础组成\r\n单 SM 包含多组 Sub Core、Warp 调度器、CUDA Core、Tensor Core，每 Block 固定调度至单一 SM 执行，不可跨 SM。\r\n#### 2. CUDA Core 与 Tensor Core 区分\r\n- CUDA Core：GPU 中负责通用计算的执行单元，支持多种运算类型；\r\n- Tensor Core 专用矩阵乘硬件，一次指令完成批量矩阵计算。\r\n#### 3. 线程数工程规范\r\n单 Block 推荐 128 \u002F 256 个线程（4 个 Warp），匹配 SM 内调度单元，充分利用硬件算力；线程过多会出现寄存器溢出，严重降低性能。\r\n\r\n## CUDA 编程模型基础\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LjFa2S9w8G7bTJvh30MJ5fayBDg%3D)\r\n\r\n### 三类函数修饰符\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4ajPVoHy6Meop76GoBMLuKZt8Vw%3D)\r\n\r\n1. `__global__`：CPU 调用、GPU 执行，标准 Kernel 函数，无返回值；\r\n2. `__device__`：仅 GPU 内部调用，建议强制内联消除函数调用开销；\r\n3. `__host__`：普通 CPU 串行函数。\r\n\r\n### Grid \u002F Block \u002F Thread 任务划分\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sOciKgLSrHZcOmcqzMCs3KItxM0%3D)\r\n\r\n- Grid：任务总容器，由若干 Block 组成；\r\n- Block：CUDA 编程中的线程组织单位，也是资源分配和调度的基本单位，同一 Block 内线程可以共享 Shared Memory；\r\n- Thread：最小执行单元，通过 BlockIdx、ThreadIdx 计算全局数据索引。\r\n\r\n### Kernel 完整执行流程\r\n1. `cudaMalloc` 在显存分配空间；\r\n2. `cudaMemcpy` 主机 ↔ 显存拷贝数据；\r\n3. 启动 Kernel 执行并行计算；\r\n4. 拷贝计算结果回主机，释放显存。\r\n\r\n### SIMT 架构特点\r\n所有线程执行同一套代码，通过不同线程索引访问不同数据；同一 Warp 内分支判断会产生线程空转（Warp 发散），尽量避免复杂控制流。\r\n\r\n##  Vector Add 向量加法\r\n### 任务特性\r\n向量加法属于逐元素无关计算，各线程处理独立数据，无线程间数据依赖，无同步需求。\r\n### 并行划分思路\r\n- 按线程均分向量元素，每个线程仅计算一组 X + Y = Z；\r\n- 线程处理数据保持内存连续，优化显存读取效率。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9jK6yhnFIeTPIjC2SgmlUihetS8%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=M8kd3ytnPHcjBKa3gbIq6sJOgps%3D)\r\n\r\n\r\n## Vector Reduction 向量规约\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7VJ6JHbyyOAN2yC8HRaVbkyXOR4%3D)\r\n\r\n### 串行 vs 并行规约\r\n- 串行规约单线程循环累加，速度极慢；\r\n- 并行规约拆分多线程分段求和，再逐层聚合得到最终结果。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TEoNwYziHNqof3HkrL20SxvEig8%3D)\r\n\r\n### 两层分层规约流程\r\n1. 线程内规约：单线程循环累加分段内全部元素，结果存入寄存器；\r\n2. Block 内树状规约：将寄存器结果写入 Shared Memory，通过 `__syncthreads()` 同步，对半聚合逐层合并。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qbyFmwHxBmriRNXAjJUd7LU0oMU%3D)\r\n\r\n### 规约加速\r\n1. Warp Shuffle：`__shfl_down_sync` 寄存器内直接交换数据，无需 Shared Memory；\r\n2. CUB 是 NVIDIA 提供的 CUDA 高性能并行算法库，工程开发优先复用。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OqqHu4JbRXN7fcH6lO85Pz2Glqw%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw14.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=W8yUnnpWqzE2akc7min50vy%2B0i4%3D)\r\n\r\n\r\n\r\n\r\n## 加入训练营\r\n#### 📺完整课程内容，请观看直播或查看课程回放：\r\n**InfiniTensor 官网**： https:\u002F\u002Fwww.infinitensor.com\r\n\r\n**B站直播**： InfiniTensor 官方直播间\r\n\r\n**视频号直播**： InfiniTensor 视频号预约\r\n\r\n**答疑交流**： 训练营官方社群⬇️\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RCGzL2mm8Hipb2jgTTOzg7DOtUs%3D)\r\n\r\n## 关注与交流\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ne2IbUxGKCs8rwPiAjr5B%2BpodTI%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-79","study-79","2026 夏季 InfiniTensor 训练营推理引擎 3：《从大模型推理到 AI 对话》",29,"\u002Fstudy-material\u002Fstudy-80","study-80","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-81","study-81","2026 夏季 InfiniTensor 训练营 CPU 并行编程 2：《CPU 指令集加速实战》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw01.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fkVCF%2Bq28sfHht31UdWLnD7Qk8I%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 27 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第四课《CUDA 基础入门：向量规约》开讲。\u003Cbr>前三课依次讲解 CPU \u002F GPU 架构差异、SIMT 基础概念、 Roofline 性能模型、GPU 多层内存架构、Shared Memory Bank Conflict，以及原子操作、Warp \u002F 树状规约、Cooperative Groups 等多级并行规约优化方案。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课从 CPU \u002F GPU 硬件差异切入，讲解 CUDA 完整编程模型、向量基础算子与并行向量规约实现，打通 CUDA 底层硬件到并行代码落地全链路。\u003C\u002Fp>\n\u003Ch2 id=\"cpu-与-gpu-硬件核心差异\">\u003Ca class=\"anchor\" href=\"#cpu-与-gpu-硬件核心差异\" aria-label=\"跳转到标题：CPU 与 GPU 硬件核心差异\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CPU 与 GPU 硬件核心差异\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MhdJ9diLJonAqMIrSfnvCRq1PXE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>CPU：大量缓存、控制单元，计算单元少，适配复杂分支逻辑；\u003C\u002Fli>\n\u003Cli>GPU 海量计算单元，缓存占比极低，面向大批量连续并行数据运算。\u003C\u002Fli>\n\u003Cli>关键区别：CPU 依赖 L3 缓存复用重复数据，GPU 依靠 Warp 切换掩盖访存延迟，不采用类似 CPU 的大容量共享 L3 Cache 设计。\u003C\u002Fli>\n\u003Cli>Warp 由 32 个线程组成，GPU 为每个 Warp 独立分配寄存器；GPU 一个 SM 可以同时驻留多个 Warp，Warp Scheduler 会在每个周期选择满足执行条件的 Warp 发射指令。当某个 Warp 因访存等原因等待时，调度器可以切换执行其他 ready Warp，从而隐藏访存延迟。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"gpu-完整存储层级与性能差距\">\u003Ca class=\"anchor\" href=\"#gpu-完整存储层级与性能差距\" aria-label=\"跳转到标题：GPU 完整存储层级与性能差距\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GPU 完整存储层级与性能差距\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw03.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=XSz4ISn6QAnjQViUKJcQ22n0kiU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"gpu-存储层级-速度由快至慢\">\u003Ca class=\"anchor\" href=\"#gpu-存储层级-速度由快至慢\" aria-label=\"跳转到标题：GPU 存储层级（速度由快至慢）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GPU 存储层级（速度由快至慢）\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>寄存器（RMEM）\u003C\u002Fstrong>：线程私有，常见单线程上限 255 个，变量直接映射寄存器；\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Shared Memory（SMEM）\u003C\u002Fstrong>：单 Block 线程共享，可编程，存在 Bug Conflict 性能隐患；\u003C\u002Fli>\n\u003Cli>显存（GMEM）：全局大容量存储。在大模型推理尤其是 Decode 阶段，显存带宽通常是影响性能的重要指标之一。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>补充：L2 Cache 硬件自动缓存，不可手动控制，极限优化场景少量利用。\u003C\u002Fp>\n\u003Ch3 id=\"异步访存硬件革新\">\u003Ca class=\"anchor\" href=\"#异步访存硬件革新\" aria-label=\"跳转到标题：异步访存硬件革新\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>异步访存硬件革新\u003C\u002Fh3>\u003Cp>GPU 支持显存 ↔ Shared Memory 异步传输，访存与计算可并行执行，大幅掩盖内存延迟；新一代 TMA 硬件进一步节省寄存器开销。\u003C\u002Fp>\n\u003Ch2 id=\"gpu-的计算单元\">\u003Ca class=\"anchor\" href=\"#gpu-的计算单元\" aria-label=\"跳转到标题：GPU 的计算单元\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>GPU 的计算单元\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw04.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ydvma2%2B5IR%2Bn4IRLeGQhEkfGWMk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"1-sm-基础组成\">\u003Ca class=\"anchor\" href=\"#1-sm-基础组成\" aria-label=\"跳转到标题：1. SM 基础组成\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. SM 基础组成\u003C\u002Fh4>\u003Cp>单 SM 包含多组 Sub Core、Warp 调度器、CUDA Core、Tensor Core，每 Block 固定调度至单一 SM 执行，不可跨 SM。\u003C\u002Fp>\n\u003Ch4 id=\"2-cuda-core-与-tensor-core-区分\">\u003Ca class=\"anchor\" href=\"#2-cuda-core-与-tensor-core-区分\" aria-label=\"跳转到标题：2. CUDA Core 与 Tensor Core 区分\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. CUDA Core 与 Tensor Core 区分\u003C\u002Fh4>\u003Cul>\n\u003Cli>CUDA Core：GPU 中负责通用计算的执行单元，支持多种运算类型；\u003C\u002Fli>\n\u003Cli>Tensor Core 专用矩阵乘硬件，一次指令完成批量矩阵计算。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-线程数工程规范\">\u003Ca class=\"anchor\" href=\"#3-线程数工程规范\" aria-label=\"跳转到标题：3. 线程数工程规范\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 线程数工程规范\u003C\u002Fh4>\u003Cp>单 Block 推荐 128 \u002F 256 个线程（4 个 Warp），匹配 SM 内调度单元，充分利用硬件算力；线程过多会出现寄存器溢出，严重降低性能。\u003C\u002Fp>\n\u003Ch2 id=\"cuda-编程模型基础\">\u003Ca class=\"anchor\" href=\"#cuda-编程模型基础\" aria-label=\"跳转到标题：CUDA 编程模型基础\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA 编程模型基础\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw05.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=LjFa2S9w8G7bTJvh30MJ5fayBDg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"三类函数修饰符\">\u003Ca class=\"anchor\" href=\"#三类函数修饰符\" aria-label=\"跳转到标题：三类函数修饰符\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>三类函数修饰符\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw06.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4ajPVoHy6Meop76GoBMLuKZt8Vw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Ccode>__global__\u003C\u002Fcode>：CPU 调用、GPU 执行，标准 Kernel 函数，无返回值；\u003C\u002Fli>\n\u003Cli>\u003Ccode>__device__\u003C\u002Fcode>：仅 GPU 内部调用，建议强制内联消除函数调用开销；\u003C\u002Fli>\n\u003Cli>\u003Ccode>__host__\u003C\u002Fcode>：普通 CPU 串行函数。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"grid-block-thread-任务划分\">\u003Ca class=\"anchor\" href=\"#grid-block-thread-任务划分\" aria-label=\"跳转到标题：Grid \u002F Block \u002F Thread 任务划分\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Grid \u002F Block \u002F Thread 任务划分\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw07.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sOciKgLSrHZcOmcqzMCs3KItxM0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Grid：任务总容器，由若干 Block 组成；\u003C\u002Fli>\n\u003Cli>Block：CUDA 编程中的线程组织单位，也是资源分配和调度的基本单位，同一 Block 内线程可以共享 Shared Memory；\u003C\u002Fli>\n\u003Cli>Thread：最小执行单元，通过 BlockIdx、ThreadIdx 计算全局数据索引。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"kernel-完整执行流程\">\u003Ca class=\"anchor\" href=\"#kernel-完整执行流程\" aria-label=\"跳转到标题：Kernel 完整执行流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Kernel 完整执行流程\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Ccode>cudaMalloc\u003C\u002Fcode> 在显存分配空间；\u003C\u002Fli>\n\u003Cli>\u003Ccode>cudaMemcpy\u003C\u002Fcode> 主机 ↔ 显存拷贝数据；\u003C\u002Fli>\n\u003Cli>启动 Kernel 执行并行计算；\u003C\u002Fli>\n\u003Cli>拷贝计算结果回主机，释放显存。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"simt-架构特点\">\u003Ca class=\"anchor\" href=\"#simt-架构特点\" aria-label=\"跳转到标题：SIMT 架构特点\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>SIMT 架构特点\u003C\u002Fh3>\u003Cp>所有线程执行同一套代码，通过不同线程索引访问不同数据；同一 Warp 内分支判断会产生线程空转（Warp 发散），尽量避免复杂控制流。\u003C\u002Fp>\n\u003Ch2 id=\"vector-add-向量加法\">\u003Ca class=\"anchor\" href=\"#vector-add-向量加法\" aria-label=\"跳转到标题：Vector Add 向量加法\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Vector Add 向量加法\u003C\u002Fh2>\u003Ch3 id=\"任务特性\">\u003Ca class=\"anchor\" href=\"#任务特性\" aria-label=\"跳转到标题：任务特性\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>任务特性\u003C\u002Fh3>\u003Cp>向量加法属于逐元素无关计算，各线程处理独立数据，无线程间数据依赖，无同步需求。\u003C\u002Fp>\n\u003Ch3 id=\"并行划分思路\">\u003Ca class=\"anchor\" href=\"#并行划分思路\" aria-label=\"跳转到标题：并行划分思路\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>并行划分思路\u003C\u002Fh3>\u003Cul>\n\u003Cli>按线程均分向量元素，每个线程仅计算一组 X + Y = Z；\u003C\u002Fli>\n\u003Cli>线程处理数据保持内存连续，优化显存读取效率。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw08.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9jK6yhnFIeTPIjC2SgmlUihetS8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw09.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=M8kd3ytnPHcjBKa3gbIq6sJOgps%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"vector-reduction-向量规约\">\u003Ca class=\"anchor\" href=\"#vector-reduction-向量规约\" aria-label=\"跳转到标题：Vector Reduction 向量规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Vector Reduction 向量规约\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw10.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=7VJ6JHbyyOAN2yC8HRaVbkyXOR4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"串行-vs-并行规约\">\u003Ca class=\"anchor\" href=\"#串行-vs-并行规约\" aria-label=\"跳转到标题：串行 vs 并行规约\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>串行 vs 并行规约\u003C\u002Fh3>\u003Cul>\n\u003Cli>串行规约单线程循环累加，速度极慢；\u003C\u002Fli>\n\u003Cli>并行规约拆分多线程分段求和，再逐层聚合得到最终结果。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw11.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=TEoNwYziHNqof3HkrL20SxvEig8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"两层分层规约流程\">\u003Ca class=\"anchor\" href=\"#两层分层规约流程\" aria-label=\"跳转到标题：两层分层规约流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>两层分层规约流程\u003C\u002Fh3>\u003Col>\n\u003Cli>线程内规约：单线程循环累加分段内全部元素，结果存入寄存器；\u003C\u002Fli>\n\u003Cli>Block 内树状规约：将寄存器结果写入 Shared Memory，通过 \u003Ccode>__syncthreads()\u003C\u002Fcode> 同步，对半聚合逐层合并。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw12.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qbyFmwHxBmriRNXAjJUd7LU0oMU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"规约加速\">\u003Ca class=\"anchor\" href=\"#规约加速\" aria-label=\"跳转到标题：规约加速\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>规约加速\u003C\u002Fh3>\u003Col>\n\u003Cli>Warp Shuffle：\u003Ccode>__shfl_down_sync\u003C\u002Fcode> 寄存器内直接交换数据，无需 Shared Memory；\u003C\u002Fli>\n\u003Cli>CUB 是 NVIDIA 提供的 CUDA 高性能并行算法库，工程开发优先复用。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw13.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OqqHu4JbRXN7fcH6lO85Pz2Glqw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_80_zw14.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=W8yUnnpWqzE2akc7min50vy%2B0i4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"加入训练营\">\u003Ca class=\"anchor\" href=\"#加入训练营\" aria-label=\"跳转到标题：加入训练营\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加入训练营\u003C\u002Fh2>\u003Ch4 id=\"完整课程内容-请观看直播或查看课程回放\">\u003Ca class=\"anchor\" href=\"#完整课程内容-请观看直播或查看课程回放\" aria-label=\"跳转到标题：📺完整课程内容，请观看直播或查看课程回放：\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>📺完整课程内容，请观看直播或查看课程回放：\u003C\u002Fh4>\u003Cp>\u003Cstrong>InfiniTensor 官网\u003C\u002Fstrong>： \u003Ca href=\"https:\u002F\u002Fwww.infinitensor.com\">https:\u002F\u002Fwww.infinitensor.com\u003C\u002Fa>\u003C\u002Fp>\n\u003Cp>\u003Cstrong>B站直播\u003C\u002Fstrong>： InfiniTensor 官方直播间\u003C\u002Fp>\n\u003Cp>\u003Cstrong>视频号直播\u003C\u002Fstrong>： InfiniTensor 视频号预约\u003C\u002Fp>\n\u003Cp>\u003Cstrong>答疑交流\u003C\u002Fstrong>： 训练营官方社群⬇️\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002F2026-xjq.png?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RCGzL2mm8Hipb2jgTTOzg7DOtUs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"关注与交流\">\u003Ca class=\"anchor\" href=\"#关注与交流\" aria-label=\"跳转到标题：关注与交流\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>关注与交流\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002Fportal_news\u002Fnews_ewm02.jpg?Expires=1792752590&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ne2IbUxGKCs8rwPiAjr5B%2BpodTI%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,134,137,139,142,146,149,152,155,157,160,163,166,169,171,173,176,179,181,183,185,188],{"id":129,"text":130,"level":110},"cpu-与-gpu-硬件核心差异","CPU 与 GPU 硬件核心差异",{"id":132,"text":133,"level":110},"gpu-完整存储层级与性能差距","GPU 完整存储层级与性能差距",{"id":135,"text":136,"level":9},"gpu-存储层级-速度由快至慢","GPU 存储层级（速度由快至慢）",{"id":138,"text":138,"level":9},"异步访存硬件革新",{"id":140,"text":141,"level":110},"gpu-的计算单元","GPU 的计算单元",{"id":143,"text":144,"level":145},"1-sm-基础组成","1. SM 基础组成",4,{"id":147,"text":148,"level":145},"2-cuda-core-与-tensor-core-区分","2. CUDA Core 与 Tensor Core 区分",{"id":150,"text":151,"level":145},"3-线程数工程规范","3. 线程数工程规范",{"id":153,"text":154,"level":110},"cuda-编程模型基础","CUDA 编程模型基础",{"id":156,"text":156,"level":9},"三类函数修饰符",{"id":158,"text":159,"level":9},"grid-block-thread-任务划分","Grid \u002F Block \u002F Thread 任务划分",{"id":161,"text":162,"level":9},"kernel-完整执行流程","Kernel 完整执行流程",{"id":164,"text":165,"level":9},"simt-架构特点","SIMT 架构特点",{"id":167,"text":168,"level":110},"vector-add-向量加法","Vector Add 向量加法",{"id":170,"text":170,"level":9},"任务特性",{"id":172,"text":172,"level":9},"并行划分思路",{"id":174,"text":175,"level":110},"vector-reduction-向量规约","Vector Reduction 向量规约",{"id":177,"text":178,"level":9},"串行-vs-并行规约","串行 vs 并行规约",{"id":180,"text":180,"level":9},"两层分层规约流程",{"id":182,"text":182,"level":9},"规约加速",{"id":184,"text":184,"level":110},"加入训练营",{"id":186,"text":187,"level":145},"完整课程内容-请观看直播或查看课程回放","📺完整课程内容，请观看直播或查看课程回放：",{"id":189,"text":189,"level":110},"关注与交流"]