[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-70":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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Loo9mNcAiqrEJeoxKJGChNjsIP0%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=aQ7DEIPmeZ6G2GUnnO42fZDzkoY%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rRUH7HCePjJ0x1Z1esoWYs0tjmE%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_70_fm.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BkL1Ck7x0%2BifrOsjbYqgaGXeyxk%3D","CUDA 编程,Roofline 模型,向量化优化,Nsight Compute,半精度计算","讲解内存墙与 Roofline 性能模型，使用 NCU 工具定位访存瓶颈，实战向量化访存、内存合并对齐，结合 FP16 半精度 SIMD 实现 GPU Kernel 提速。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 CUDA 编程 2：《性能模型与逐元素优化》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ina36WAf%2BdI6WwTSwIRtw2hxBx4%3D)\r\n\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 21 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第二课《性能模型与逐元素优化》开讲。\r\n前置课程《并行编程导论与 CUDA 入门》详细讲解了 CUDA 编程基础、线程层级结构与向量加法示例。但也发现了一个残酷的现实：**数据传输耗时远超计算本身**。\u003C\u002Fspan>\r\n\r\n本节课我们将深入剖析性能瓶颈的根源，掌握量化分析工具，并通过“向量化”这一核心优化手段，带你突破内存墙限制，实现性能的提升。\r\n\r\n## 温故知新：从现象到本质\r\n- **上节回顾**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=R9Qazo1DgsDwUHVun3u6xZrDiXo%3D)\r\n\r\n- **性能痛点**：为何 GPU 满载时，H2D\u002FD2H 传输耗时仍远大于 Kernel 计算？\r\n\r\n- **核心矛盾**：CPU-GPU 通信带宽 vs GPU 内部算力，谁是真正的瓶颈？\r\n\r\n## 内存墙与 Roofline 模型\r\n### 理解“内存墙” (Memory Wall)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2B6AtYUyaUNpBo23FSCbYkLY72tk%3D)\r\n\r\n- **摩尔定律的失衡**：处理器性能增速 vs 内存带宽增速的巨大 Gap\r\n- **冯·诺依曼架构局限**：存储与计算分离导致的频繁数据搬运\r\n- **通信**：PCIe\u002FNVLink 带宽增长滞后于算力增长\r\n\r\n### Roofline 模型详解\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ECvpl6ZeiQYZWeqiogssLI%2BtoWg%3D)\r\n\r\n1. **核心坐标**：计算强度 (Operational Intensity, FLOPS\u002FByte) vs 性能 (FLOPS)\r\n\r\n2. **两大区域**：\r\n   - 🔴 Memory Bound（访存密集型）：受限于内存带宽 (β × I)\r\n   - 🟢 Compute Bound（计算密集型）：受限于峰值算力 (π)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y8kT06rBqQdxjBzaF3tDiZt%2FvMg%3D)\r\n\r\n3. **关键拐点 (Ridge Point)**：达到硬件理论最高性能所需的最低计算强度 ($I_{max} = pi \u002F \beta$)\r\n\r\n4. **实战分析**：以 A100 为例，量化分析 FP32 向量加法为何是典型的访存密集型任务\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cd5njnyTadPCclhqfENPTf%2B73l0%3D)\r\n\r\n\r\n## Nsight Compute (NCU)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WHaD4nAM4InrUkav%2Buz9iZL1JXI%3D)\r\n\r\n1. **工具定位**：区别于 Nsight Systems，NCU 专注于 Kernel 级微观性能分析\r\n2. **核心功能**：\r\n   - Summary 页：快速查看吞吐率与运行时间\r\n   - Details 页：GPU Speed of Light Throughput 与 Roofline 图表自动生成\r\n3. **实测验证**：通过 NCU 确认向量加法的内存利用率高达 70%，验证访存瓶颈假设\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xKIFPOInk0%2B0hVCTzwPRYygcH6c%3D)\r\n\r\n\r\n## 向量化访存 (Vectorization)\r\n\r\n### 什么是向量化？\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=POIkTRQEGSUp8rw7MVF62uyWcTw%3D)\r\n\r\n- **SIMD vs SIMT**：单指令多数据流与单指令多线程的协同而非冲突\r\n- **核心思想**：单次指令读写多个标量数据，提升内存带宽利用率\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GaT4hXePtwKrpvy%2FZxBKUi5JJf0%3D)\r\n\r\n\r\n### CUDA 内置向量类型实战\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JAuEd%2B56UdlyyIX8Qi1TCg8p5KI%3D)\r\n\r\n- **数据类型**：`float2`、`float4` 等内置向量化访存类型\r\n- **代码重构**：使用模板函数 + `__device__` 封装通用加法逻辑\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kPGj2Dn6aE982jTiFz3xP24hIZE%3D)\r\n\r\n\r\n- **性能实测**：`float4` 相比 `float` 性能显著提升，内存利用率逼近 90%\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WXCnQ7%2FWDkvxaKHKmFaoLyDd4ko%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=blc%2BuIIwDtGiKK4bshWeXv8btF8%3D)\r\n\r\n\r\n### 避坑：向量化的边界\r\n1. **并非越大越好**：为何 `Vec8` 反而比 `Vec4` 慢？（寄存器溢出 Register Spilling）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JyV4KUpPxcJy9oDzmJReRCcka5w%3D)\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ybi%2BpdCU4j60a%2FvrS3ymzBxHw60%3D)\r\n\r\n2. **`float3` 为何比 `float` 慢？**：\r\n   - **内存对齐 (Memory Alignment)**：非 2 的幂次导致无法对齐内存事务边界\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qNM3Ej9GCXjmiJjD0QPrFF4JC9U%3D)\r\n\r\n\r\n   - **内存合并 (Memory Coalescing)**：Warp 内连续访问 + 对齐 = 高效合并\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iFrX6r66or0CajAhZLm6IXnrWQM%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rbD%2F3whSGBctIQj7i95JyvETSmw%3D)\r\n\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uE2%2BQOeVVGDel7NAKiyKUDk1RdI%3D)\r\n\r\n\r\n## 进阶优化：半精度计算与 SIMD 指令\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BTjRUa1H4zlIsrt76Y4YUQwUvQM%3D)\r\n\r\n### 半精度（`half`）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9evE6iyYTrpWnqrpCppHio0toCA%3D)\r\n\r\n\r\n1. **降低访存压力**：FP16 (`half`) 数据类型原理与优势\r\n\r\n2. 半精度并行加法示例：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw23.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0JjsEW29Q6VzJrE1Cb52IlJS8Qs%3D)\r\n\r\n   - 头文件与类型转换 (`__float2half`, `__half2float`)\r\n   - 舍入模式解析 (RD\u002FRU\u002FRN\u002FRZ)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw24.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OvPqQkrZ%2FkE9pHOxsdXGtEHzVlg%3D)\r\n\r\n3. **向量化计算**：使用 `__hadd2` 实现真正的 SIMD 半精度加法\r\n\r\n4. **Roofline 再分析**：FP16 如何同时实现“右移”（提高计算强度）与“上移”（利用 Tensor Core 提升天花板）\r\n\r\n\r\n**❓️课后思考题**：\r\n1. 若必须使用 `float3` 结构，如何优化其内存访问性能？\r\n2. 除内置向量类型外，还有哪些方式可实现向量化访存？（提示：与思考题1相关）\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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wGzqtB9HsaU4IHS5eDT0qyV2lsg%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9Am%2FQfw7y%2BpppXGy6g1VG29SISM%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-69","study-69","2026 夏季 InfiniTensor 训练营前沿模型 2：《检索增强的语言模型》",39,"\u002Fstudy-material\u002Fstudy-70","study-70","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-71","study-71","2026 夏季 InfiniTensor 训练营 OpenCL 编程 1：《OpenCL 概述与运行时》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Ina36WAf%2BdI6WwTSwIRtw2hxBx4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 21 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第二课《性能模型与逐元素优化》开讲。\u003Cbr>前置课程《并行编程导论与 CUDA 入门》详细讲解了 CUDA 编程基础、线程层级结构与向量加法示例。但也发现了一个残酷的现实：\u003Cstrong>数据传输耗时远超计算本身\u003C\u002Fstrong>。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课我们将深入剖析性能瓶颈的根源，掌握量化分析工具，并通过“向量化”这一核心优化手段，带你突破内存墙限制，实现性能的提升。\u003C\u002Fp>\n\u003Ch2 id=\"温故知新-从现象到本质\">\u003Ca class=\"anchor\" href=\"#温故知新-从现象到本质\" aria-label=\"跳转到标题：温故知新：从现象到本质\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>温故知新：从现象到本质\u003C\u002Fh2>\u003Cul>\n\u003Cli>\u003Cstrong>上节回顾\u003C\u002Fstrong>：\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=R9Qazo1DgsDwUHVun3u6xZrDiXo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cp>\u003Cstrong>性能痛点\u003C\u002Fstrong>：为何 GPU 满载时，H2D\u002FD2H 传输耗时仍远大于 Kernel 计算？\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>核心矛盾\u003C\u002Fstrong>：CPU-GPU 通信带宽 vs GPU 内部算力，谁是真正的瓶颈？\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"内存墙与-roofline-模型\">\u003Ca class=\"anchor\" href=\"#内存墙与-roofline-模型\" aria-label=\"跳转到标题：内存墙与 Roofline 模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>内存墙与 Roofline 模型\u003C\u002Fh2>\u003Ch3 id=\"理解-内存墙-memory-wall\">\u003Ca class=\"anchor\" href=\"#理解-内存墙-memory-wall\" aria-label=\"跳转到标题：理解“内存墙” (Memory Wall)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>理解“内存墙” (Memory Wall)\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2B6AtYUyaUNpBo23FSCbYkLY72tk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>摩尔定律的失衡\u003C\u002Fstrong>：处理器性能增速 vs 内存带宽增速的巨大 Gap\u003C\u002Fli>\n\u003Cli>\u003Cstrong>冯·诺依曼架构局限\u003C\u002Fstrong>：存储与计算分离导致的频繁数据搬运\u003C\u002Fli>\n\u003Cli>\u003Cstrong>通信\u003C\u002Fstrong>：PCIe\u002FNVLink 带宽增长滞后于算力增长\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"roofline-模型详解\">\u003Ca class=\"anchor\" href=\"#roofline-模型详解\" aria-label=\"跳转到标题：Roofline 模型详解\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Roofline 模型详解\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ECvpl6ZeiQYZWeqiogssLI%2BtoWg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cp>\u003Cstrong>核心坐标\u003C\u002Fstrong>：计算强度 (Operational Intensity, FLOPS\u002FByte) vs 性能 (FLOPS)\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>两大区域\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>🔴 Memory Bound（访存密集型）：受限于内存带宽 (β × I)\u003C\u002Fli>\n\u003Cli>🟢 Compute Bound（计算密集型）：受限于峰值算力 (π)\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=y8kT06rBqQdxjBzaF3tDiZt%2FvMg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>\u003Cp>\u003Cstrong>关键拐点 (Ridge Point)\u003C\u002Fstrong>：达到硬件理论最高性能所需的最低计算强度 ($I_{max} = pi \u002F \beta$)\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>实战分析\u003C\u002Fstrong>：以 A100 为例，量化分析 FP32 向量加法为何是典型的访存密集型任务\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=cd5njnyTadPCclhqfENPTf%2B73l0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"nsight-compute-ncu\">\u003Ca class=\"anchor\" href=\"#nsight-compute-ncu\" aria-label=\"跳转到标题：Nsight Compute (NCU)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Nsight Compute (NCU)\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WHaD4nAM4InrUkav%2Buz9iZL1JXI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>工具定位\u003C\u002Fstrong>：区别于 Nsight Systems，NCU 专注于 Kernel 级微观性能分析\u003C\u002Fli>\n\u003Cli>\u003Cstrong>核心功能\u003C\u002Fstrong>：\u003Cul>\n\u003Cli>Summary 页：快速查看吞吐率与运行时间\u003C\u002Fli>\n\u003Cli>Details 页：GPU Speed of Light Throughput 与 Roofline 图表自动生成\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003Cli>\u003Cstrong>实测验证\u003C\u002Fstrong>：通过 NCU 确认向量加法的内存利用率高达 70%，验证访存瓶颈假设\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=xKIFPOInk0%2B0hVCTzwPRYygcH6c%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"向量化访存-vectorization\">\u003Ca class=\"anchor\" href=\"#向量化访存-vectorization\" aria-label=\"跳转到标题：向量化访存 (Vectorization)\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>向量化访存 (Vectorization)\u003C\u002Fh2>\u003Ch3 id=\"什么是向量化\">\u003Ca class=\"anchor\" href=\"#什么是向量化\" aria-label=\"跳转到标题：什么是向量化？\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>什么是向量化？\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=POIkTRQEGSUp8rw7MVF62uyWcTw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>SIMD vs SIMT\u003C\u002Fstrong>：单指令多数据流与单指令多线程的协同而非冲突\u003C\u002Fli>\n\u003Cli>\u003Cstrong>核心思想\u003C\u002Fstrong>：单次指令读写多个标量数据，提升内存带宽利用率\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=GaT4hXePtwKrpvy%2FZxBKUi5JJf0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"cuda-内置向量类型实战\">\u003Ca class=\"anchor\" href=\"#cuda-内置向量类型实战\" aria-label=\"跳转到标题：CUDA 内置向量类型实战\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA 内置向量类型实战\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JAuEd%2B56UdlyyIX8Qi1TCg8p5KI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>数据类型\u003C\u002Fstrong>：\u003Ccode>float2\u003C\u002Fcode>、\u003Ccode>float4\u003C\u002Fcode> 等内置向量化访存类型\u003C\u002Fli>\n\u003Cli>\u003Cstrong>代码重构\u003C\u002Fstrong>：使用模板函数 + \u003Ccode>__device__\u003C\u002Fcode> 封装通用加法逻辑\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=kPGj2Dn6aE982jTiFz3xP24hIZE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>性能实测\u003C\u002Fstrong>：\u003Ccode>float4\u003C\u002Fcode> 相比 \u003Ccode>float\u003C\u002Fcode> 性能显著提升，内存利用率逼近 90%\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WXCnQ7%2FWDkvxaKHKmFaoLyDd4ko%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=blc%2BuIIwDtGiKK4bshWeXv8btF8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"避坑-向量化的边界\">\u003Ca class=\"anchor\" href=\"#避坑-向量化的边界\" aria-label=\"跳转到标题：避坑：向量化的边界\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>避坑：向量化的边界\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Cstrong>并非越大越好\u003C\u002Fstrong>：为何 \u003Ccode>Vec8\u003C\u002Fcode> 反而比 \u003Ccode>Vec4\u003C\u002Fcode> 慢？（寄存器溢出 Register Spilling）\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=JyV4KUpPxcJy9oDzmJReRCcka5w%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ybi%2BpdCU4j60a%2FvrS3ymzBxHw60%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"2\">\n\u003Cli>\u003Cstrong>\u003Ccode>float3\u003C\u002Fcode> 为何比 \u003Ccode>float\u003C\u002Fcode> 慢？\u003C\u002Fstrong>：\u003Cul>\n\u003Cli>\u003Cstrong>内存对齐 (Memory Alignment)\u003C\u002Fstrong>：非 2 的幂次导致无法对齐内存事务边界\u003C\u002Fli>\n\u003C\u002Ful>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qNM3Ej9GCXjmiJjD0QPrFF4JC9U%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>内存合并 (Memory Coalescing)\u003C\u002Fstrong>：Warp 内连续访问 + 对齐 = 高效合并\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=iFrX6r66or0CajAhZLm6IXnrWQM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=rbD%2F3whSGBctIQj7i95JyvETSmw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uE2%2BQOeVVGDel7NAKiyKUDk1RdI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"进阶优化-半精度计算与-simd-指令\">\u003Ca class=\"anchor\" href=\"#进阶优化-半精度计算与-simd-指令\" aria-label=\"跳转到标题：进阶优化：半精度计算与 SIMD 指令\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>进阶优化：半精度计算与 SIMD 指令\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=BTjRUa1H4zlIsrt76Y4YUQwUvQM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"半精度-half\">\u003Ca class=\"anchor\" href=\"#半精度-half\" aria-label=\"跳转到标题：半精度（half）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>半精度（\u003Ccode>half\u003C\u002Fcode>）\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9evE6iyYTrpWnqrpCppHio0toCA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cp>\u003Cstrong>降低访存压力\u003C\u002Fstrong>：FP16 (\u003Ccode>half\u003C\u002Fcode>) 数据类型原理与优势\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>半精度并行加法示例：\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw23.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0JjsEW29Q6VzJrE1Cb52IlJS8Qs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>头文件与类型转换 (\u003Ccode>__float2half\u003C\u002Fcode>, \u003Ccode>__half2float\u003C\u002Fcode>)\u003C\u002Fli>\n\u003Cli>舍入模式解析 (RD\u002FRU\u002FRN\u002FRZ)\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_70_zw24.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OvPqQkrZ%2FkE9pHOxsdXGtEHzVlg%3D\" alt=\"\">\u003C\u002Fp>\n\u003Col start=\"3\">\n\u003Cli>\u003Cp>\u003Cstrong>向量化计算\u003C\u002Fstrong>：使用 \u003Ccode>__hadd2\u003C\u002Fcode> 实现真正的 SIMD 半精度加法\u003C\u002Fp>\n\u003C\u002Fli>\n\u003Cli>\u003Cp>\u003Cstrong>Roofline 再分析\u003C\u002Fstrong>：FP16 如何同时实现“右移”（提高计算强度）与“上移”（利用 Tensor Core 提升天花板）\u003C\u002Fp>\n\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cstrong>❓️课后思考题\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Col>\n\u003Cli>若必须使用 \u003Ccode>float3\u003C\u002Fcode> 结构，如何优化其内存访问性能？\u003C\u002Fli>\n\u003Cli>除内置向量类型外，还有哪些方式可实现向量化访存？（提示：与思考题1相关）\u003C\u002Fli>\n\u003C\u002Fol>\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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=wGzqtB9HsaU4IHS5eDT0qyV2lsg%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=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9Am%2FQfw7y%2BpppXGy6g1VG29SISM%3D\" alt=\"\">\u003C\u002Fp>\n",[128,131,134,137,140,143,146,149,152,155,158,161,163,167],{"id":129,"text":130,"level":110},"温故知新-从现象到本质","温故知新：从现象到本质",{"id":132,"text":133,"level":110},"内存墙与-roofline-模型","内存墙与 Roofline 模型",{"id":135,"text":136,"level":9},"理解-内存墙-memory-wall","理解“内存墙” (Memory Wall)",{"id":138,"text":139,"level":9},"roofline-模型详解","Roofline 模型详解",{"id":141,"text":142,"level":110},"nsight-compute-ncu","Nsight Compute (NCU)",{"id":144,"text":145,"level":110},"向量化访存-vectorization","向量化访存 (Vectorization)",{"id":147,"text":148,"level":9},"什么是向量化","什么是向量化？",{"id":150,"text":151,"level":9},"cuda-内置向量类型实战","CUDA 内置向量类型实战",{"id":153,"text":154,"level":9},"避坑-向量化的边界","避坑：向量化的边界",{"id":156,"text":157,"level":110},"进阶优化-半精度计算与-simd-指令","进阶优化：半精度计算与 SIMD 指令",{"id":159,"text":160,"level":9},"半精度-half","半精度（half）",{"id":162,"text":162,"level":110},"加入训练营",{"id":164,"text":165,"level":166},"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：",4,{"id":168,"text":168,"level":110},"关注与交流"]