[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-98":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=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Fy5fJ3AB6DC%2F4ixX2oU%2Fh4v9%2F%2Bc%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=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Gsep2R2fMub4ugC6MyVcAmFrb%2FE%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=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=J07ePwwPpIjasanoQ4%2BRXZYgbnY%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_98_fm.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PcI37p%2FOV7sJX71YPEg9b8kN46A%3D","CUDA,模型量化,PTX 汇编,算子调优,Tensor Core","2026 夏季 InfiniTensor 训练营《量化与工业级算子调优部署》讲解 FP 16 \u002F BF 16 等全系列低精度数据类型；完整拆解量化四大分类与 INT 8 量化落地流程；深入 PTX 底层汇编、编译器修饰符、L1 缓存控制、Warp 线程发散优化；介绍 ILP \u002F TLP 并行平衡、SOA 数据排布，最后梳理 cuDNN \u002F cuBLAS \u002F CUB 官方加速库。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 CUDA 编程 9：《量化与工业级算子调优部署》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw01.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RO93Voio%2FK4OUTniRZynceLLmbU%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">8 月 10 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第九课《量化与工业级算子调优部署》开讲。\r\n上节课《异步并行、底层控制与系统优化》聚焦系统层面性能优化，讲解 NCU 瓶颈诊断、双缓冲存算重叠、异步流与统一内存、CUDA Graph、NCCL 多卡通信及阿姆达尔\u002F古斯塔夫优化定律等内容。\u003C\u002Fspan>\r\n\r\n本节课作为 CUDA 编程系列收尾课程，介绍各类浮点\u002F整数低精度类型、量化全体系、PTX 底层指令、编译器调优、分支发散优化与官方加速库落地部署等相关内容。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw02.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A7iMztPSPA9tKcby8GHj3a%2BPsIw%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw03.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vAzXqkmmaYdobdc3jsRQExWn%2FPk%3D)\r\n\r\n## 低精度浮点数据类型完整解析\r\n### 主流低精度类型特性对比\r\n1.**FP 16（Half）**\r\n- 5 bit 指数 + 10 bit 尾数，精度更高、动态范围偏小，通用推理基础类型。\r\n\r\n2.**BF 16（Brain Floating Point）**\r\n- 8 bit 指数 + 7 bit 尾数，与 FP 32 动态范围完全对齐，训练场景首选。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw04.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fRPKqmv0FPYcX7qmzNUx9rXxgPk%3D)\r\n\r\n3.**TF 32**\r\n- 保持 FP32 的 8 bit 指数位；\r\n- 尾数缩减至 10 bit，提高 Tensor Core 矩阵计算性能；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw05.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=na%2BLDHT5%2BjOdJTDJcgJkg%2Fgg7UI%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw06.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WLy%2FM1fZpYYNbz09N7JhdpmUbU8%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw07.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qCA0EDXT8dclQPtLPBeE83cUkjA%3D)\r\n\r\n4.**FP 8 双变体**\r\n- E4M3：4 bit 指数 + 3 bit 尾数，精度高，推理权重常用；\r\n- E5M2：5 bit 指数 + 2 bit 尾数，动态范围更大，超大数值场景适配；\r\n- Hopper 架构原生硬件支持 FP 8 混合乘加。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw08.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FatITH9G0fRTEnrBiHg35SHCuRs%3D)\r\n\r\n5.**INT 8 整数类型**\r\n- 不包含指数和尾数，通过整数编码表示数值；\r\n- 通常通过 scale \u002F zero point 将 FP32 映射到 INT8；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw09.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lmLkT9IYlZkBjZ6SliEDGT%2BTAoE%3D)\r\n\r\n6.拓展高精度 FP 128\r\n- 金融、科学计算等高敏感精度场景使用。\r\n\r\n\r\n## 深度学习量化完整技术体系\r\n### 量化基础定义\r\n- 量化 = 将高精度 FP 32 数值映射至低位宽整数\u002F浮点区间；\r\n- 正向映射为量化，反向恢复数值为反量化，核心权衡显存占用与任务精度损失。\r\n\r\n### 量化四大分类维度\r\n#### 1. 线性 vs 非线性量化\r\n- 线性：通过 scale \u002F zero 线性缩放，计算简单、推理开销低；\r\n- 非线性：对数、聚类等映射，数值分布拟合更好，但运行时计算开销更高。\r\n\r\n#### 2. 对称 vs 非对称量化\r\n- 对称：无零点偏移，正负数值区间完全对称，计算逻辑极简；\r\n- 非对称：增加 zero 偏移适配非均匀数据分布，精度上限更高。\r\n\r\n#### 3. 静态 vs 动态量化\r\n- 静态：离线标定固定 scale，推理无额外计算开销；\r\n- 动态：每批数据实时计算缩放因子，精度更优但增加运行时计算。\r\n\r\n#### 4. 量化粒度（逐层\u002F通道\u002F分组）\r\n1. 张量级：整张张量共用一组 scale，实现最简单；\r\n2. 通道级：每个通道独立缩放，精度提升；\r\n3. 组级（MXFP4 等方案）：多个连续元素共享一个缩放因子，兼顾显存与精度。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw10.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jLusNbnv%2FF89L5UayCfFibXMG5g%3D)\r\n\r\n### INT 8 量化完整执行流程\r\n1. 推理前：FP 32 权重经标定计算 scale \u002F zero，量化存储为 INT 8；\r\n2. 推理前向：根据硬件和算子实现，可采用 INT8 输入直接进行整数计算，或者在计算前转换为低精度浮点执行；\r\n3. 进阶优化：全程 INT 8 向量化访存与计算，减少 HBM 读写次数；\r\n4. 性能收益：同等硬件下吞吐相比 FP 16 可提升近一倍，代价为有限精度损耗。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw11.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g%2FjYtWUxqZQlY8kEY8vpobjr4hA%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw12.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vQ47P21BsTy2KBqRHWI8h8VqQ%2B0%3D)\r\n\r\n### 量化取舍原则\r\n- 量化依靠牺牲部分数值精度换取显存、算力收益；\r\n- 文本、图像生成任务可适度采用，金融、仿真等强精度场景不建议无限制低位宽量化。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw13.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9m%2FzTLFWpvr0BddLxnl0lXpQXTE%3D)\r\n\r\n## 性能优化——PTX\r\n### PTX 核心价值\r\n介于 C++ 与硬件指令之间的虚拟汇编层：\r\n1. 提供上层代码未暴露的细粒度硬件操作；\r\n2. 规避编译器保守优化，手动调度 MMA、异步拷贝等专用指令；\r\n3. 跨 GPU 架构向前兼容，提高代码迁移灵活性。\r\n\r\n### 典型 PTX 落地项目案例\r\n1. DeepSeek MoE 通信库：自定义分组门控 PTX 汇编优化通信开销；\r\n2. DeepGEMM FP 8 矩阵乘：利用 PTX 中的矩阵加载\u002F存储指令优化 Tensor Core 数据搬运；\r\n3. Flash LLM：Load\u002FStore 专用内存指令优化长上下文访存。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw14.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FhFdV6wizQUZPeTUjzzG5dOnLug%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw15.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Q9DRHSVzssYYlZt7404sVUFkMwM%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw16.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hiqBPj%2F72R7Is9aKusMKt3WQ8b0%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw17.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FMMDLwtD4kiRV3mbKEOj%2Fc0hn7E%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw18.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=p4ow%2BWgNNxNUKJVwFiyKb3Qtl9M%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw19.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sJPlOYS6pNJi%2BC3O%2FJLcMmBDKfE%3D)\r\n\r\n### Runtime vs Driver API\r\n- Runtime API：高层封装，开发效率高，通用场景首选；\r\n- Driver API：底层管控，支持动态加载、硬件精细调优，算子极致优化场景使用。\r\n\r\n## 微指令调优\r\n### 循环优化指令\r\n`#pragma unroll` 循环展开：消除循环分支开销、提升 ILP 指令并行；\r\n注意过度展开会造成代码膨胀、寄存器溢出，需控制展开倍数。\r\n\r\n### 断言与约束修饰符\r\n1. `__builtin_assume(cond)`：向编译器提供条件约束信息，帮助优化代码路径；\r\n2. `__align__`：指定数据对齐要求，提高满足硬件访存对齐条件的可能性。\r\n\r\n### 缓存控制编译选项\r\n1. CA（默认）：完整启用 L1 缓存；\r\n2. CG：绕过 L1，全部资源分配给共享内存；\r\n3. CS：流式内存标记，缓存数据使用后立即丢弃，减少缓存污染。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw20.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VAFYnhBnQ9yuEdERscR5T%2FctiyQ%3D)\r\n\r\n### 浮点与计算编译开关\r\n1. `--use_fast_math`：近似三角函数，提速小幅损失精度；\r\n2. `--ftz=true`：刷新非规格浮点数为 0，简化计算；\r\n3 寄存器上限限制：防止寄存器溢出至本地内存拖慢速度。\r\n\r\n## Warp 线程发散问题与优化\r\n### 发散成因\r\n同一 Warp（32 线程）内执行不同分支路径，GPU 串行执行所有分支，硬件利用率暴跌，NCU 可观测长 Stall 指标。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw21.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j3XS39uDgUg1QkUbkHdTROGUYk8%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw22.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x9IgbTB5E%2BBn%2BjkaCkOb4GKBrXQ%3D)\r\n\r\n### 主流规避方案\r\n1. Mask 掩码计算：所有线程完整执行计算，通过掩码控制写入，消除分支跳转；\r\n2. 算法重构：拆分条件逻辑，对齐线程执行路径；\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw23.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D5T%2BZtJXBvXagJQ1IrwCqzeDNj8%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw24.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SOwnFKQaCM81JsOvOazRcysAWC8%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw25.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4bPkd53B5rTx9khvQI%2FzMJkRE4Y%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw26.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qWtUTL0Y7u2fBmVtkCUswqvufCQ%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw27.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CuK2hIFNrASJur3PTPgOdxguaAk%3D)\r\n\r\n### 指标观测\r\nNCU 查看 `average predicate thread` 平均激活线程数，数值越低发散越严重。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw28.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hK46sgOKmMH6CzgTYeiyukKxNw0%3D)\r\n\r\n## 并行平衡与数据排布调优\r\n### ILP 指令级并行\r\n无依赖指令重叠执行，利用访存等待周期同步计算；手写 PTX 可手动调度指令顺序放大 ILP 收益。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw29.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uzgLV524%2FzOyJm%2FvH%2BUfRedMV54%3D)\r\n\r\n### TLP 线程级并行\r\n多线程分摊计算负载；ILP 与 TLP 需要平衡，单一维度拉满会占用过多寄存器\u002F共享内存。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw30.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HIG3uRvxmOxRwaqg6bBhuJUi9j4%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw31.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A9DuxBGX6WSU%2FD2YDv4qdEO1qIs%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw32.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4tEccuHOKzEyEE7LBvQQLXKbAf4%3D)\r\n\r\n### AOS \u002F SOA 数据结构选择\r\n- AOS：结构体数组，随机访存友好，批量连续读取效率差；\r\n- SOA：数组结构体，同字段连续排布，向量化加载、矩阵乘场景性能显著优于 AOS。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw33.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=S%2BHshDzz6MTEqTm1pmei9sIkTPk%3D)\r\n\r\n### Block Size 调优\r\n默认 256 非最优，不同算子、硬件架构适配最优 Block 尺寸，需结合寄存器占用、共享内存上限测试。\r\n\r\n## CUDA 官方库\r\n#### 1. cuDNN\r\n深度神经网络专用库，卷积、归一化、激活高度优化，原生支持 Tensor Core 混合精度、算子融合。\r\n#### 2. cuBLAS\r\n标准线性代数库，向量\u002F矩阵乘全精度实现。\r\n#### 3. CUB\r\nCUDA 高性能并行原语库，提供 Block、Warp、Device 级别的扫描、归约、排序等基础操作。\r\n#### 4. CUTLASS\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw34.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sg7njRdIe1%2BZWWQnikuUuSLz1Ko%3D)\r\n\r\n#### 5.cuSPARSE\r\n稀疏矩阵专用运算库，支持多种稀疏存储格式与分解求解。\r\n\r\n\r\n## CUDA 全系列课程完整复盘\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw35.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MFvf09VwvzDDYlZR471eKB9iUrs%3D)\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\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\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OqqoZKUCZDwCVp9tER7EkbY9quU%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=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G1NJskF7AaercUrU0ISrAy2%2F%2F6c%3D)",{"publishTime":7,"createTime":109,"canonicalPath":114,"id":115,"title":116},"\u002Fstudy-material\u002Fstudy-97","study-97","2026 夏季 InfiniTensor 训练营前沿模型 4、AI 编程与智能体应用 8：《大模型自主智能体》",12,"\u002Fstudy-material\u002Fstudy-98","study-98","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-99","study-99","2026 夏季 InfiniTensor 训练营训练系统 5：《强化学习基础与 DeepSeek 训练介绍》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw01.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RO93Voio%2FK4OUTniRZynceLLmbU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">8 月 10 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第九课《量化与工业级算子调优部署》开讲。\u003Cbr>上节课《异步并行、底层控制与系统优化》聚焦系统层面性能优化，讲解 NCU 瓶颈诊断、双缓冲存算重叠、异步流与统一内存、CUDA Graph、NCCL 多卡通信及阿姆达尔\u002F古斯塔夫优化定律等内容。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课作为 CUDA 编程系列收尾课程，介绍各类浮点\u002F整数低精度类型、量化全体系、PTX 底层指令、编译器调优、分支发散优化与官方加速库落地部署等相关内容。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw02.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A7iMztPSPA9tKcby8GHj3a%2BPsIw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw03.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vAzXqkmmaYdobdc3jsRQExWn%2FPk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"低精度浮点数据类型完整解析\">\u003Ca class=\"anchor\" href=\"#低精度浮点数据类型完整解析\" aria-label=\"跳转到标题：低精度浮点数据类型完整解析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>低精度浮点数据类型完整解析\u003C\u002Fh2>\u003Ch3 id=\"主流低精度类型特性对比\">\u003Ca class=\"anchor\" href=\"#主流低精度类型特性对比\" aria-label=\"跳转到标题：主流低精度类型特性对比\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>主流低精度类型特性对比\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>FP 16（Half）\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>5 bit 指数 + 10 bit 尾数，精度更高、动态范围偏小，通用推理基础类型。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>BF 16（Brain Floating Point）\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>8 bit 指数 + 7 bit 尾数，与 FP 32 动态范围完全对齐，训练场景首选。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw04.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=fRPKqmv0FPYcX7qmzNUx9rXxgPk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>TF 32\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>保持 FP32 的 8 bit 指数位；\u003C\u002Fli>\n\u003Cli>尾数缩减至 10 bit，提高 Tensor Core 矩阵计算性能；\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw05.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=na%2BLDHT5%2BjOdJTDJcgJkg%2Fgg7UI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw06.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=WLy%2FM1fZpYYNbz09N7JhdpmUbU8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw07.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qCA0EDXT8dclQPtLPBeE83cUkjA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>4.\u003Cstrong>FP 8 双变体\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>E4M3：4 bit 指数 + 3 bit 尾数，精度高，推理权重常用；\u003C\u002Fli>\n\u003Cli>E5M2：5 bit 指数 + 2 bit 尾数，动态范围更大，超大数值场景适配；\u003C\u002Fli>\n\u003Cli>Hopper 架构原生硬件支持 FP 8 混合乘加。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw08.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FatITH9G0fRTEnrBiHg35SHCuRs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>5.\u003Cstrong>INT 8 整数类型\u003C\u002Fstrong>\u003C\u002Fp>\n\u003Cul>\n\u003Cli>不包含指数和尾数，通过整数编码表示数值；\u003C\u002Fli>\n\u003Cli>通常通过 scale \u002F zero point 将 FP32 映射到 INT8；\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw09.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=lmLkT9IYlZkBjZ6SliEDGT%2BTAoE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>6.拓展高精度 FP 128\u003C\u002Fp>\n\u003Cul>\n\u003Cli>金融、科学计算等高敏感精度场景使用。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"深度学习量化完整技术体系\">\u003Ca class=\"anchor\" href=\"#深度学习量化完整技术体系\" aria-label=\"跳转到标题：深度学习量化完整技术体系\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>深度学习量化完整技术体系\u003C\u002Fh2>\u003Ch3 id=\"量化基础定义\">\u003Ca class=\"anchor\" href=\"#量化基础定义\" aria-label=\"跳转到标题：量化基础定义\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化基础定义\u003C\u002Fh3>\u003Cul>\n\u003Cli>量化 = 将高精度 FP 32 数值映射至低位宽整数\u002F浮点区间；\u003C\u002Fli>\n\u003Cli>正向映射为量化，反向恢复数值为反量化，核心权衡显存占用与任务精度损失。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"量化四大分类维度\">\u003Ca class=\"anchor\" href=\"#量化四大分类维度\" aria-label=\"跳转到标题：量化四大分类维度\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化四大分类维度\u003C\u002Fh3>\u003Ch4 id=\"1-线性-vs-非线性量化\">\u003Ca class=\"anchor\" href=\"#1-线性-vs-非线性量化\" aria-label=\"跳转到标题：1. 线性 vs 非线性量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. 线性 vs 非线性量化\u003C\u002Fh4>\u003Cul>\n\u003Cli>线性：通过 scale \u002F zero 线性缩放，计算简单、推理开销低；\u003C\u002Fli>\n\u003Cli>非线性：对数、聚类等映射，数值分布拟合更好，但运行时计算开销更高。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"2-对称-vs-非对称量化\">\u003Ca class=\"anchor\" href=\"#2-对称-vs-非对称量化\" aria-label=\"跳转到标题：2. 对称 vs 非对称量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. 对称 vs 非对称量化\u003C\u002Fh4>\u003Cul>\n\u003Cli>对称：无零点偏移，正负数值区间完全对称，计算逻辑极简；\u003C\u002Fli>\n\u003Cli>非对称：增加 zero 偏移适配非均匀数据分布，精度上限更高。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"3-静态-vs-动态量化\">\u003Ca class=\"anchor\" href=\"#3-静态-vs-动态量化\" aria-label=\"跳转到标题：3. 静态 vs 动态量化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. 静态 vs 动态量化\u003C\u002Fh4>\u003Cul>\n\u003Cli>静态：离线标定固定 scale，推理无额外计算开销；\u003C\u002Fli>\n\u003Cli>动态：每批数据实时计算缩放因子，精度更优但增加运行时计算。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch4 id=\"4-量化粒度-逐层-通道-分组\">\u003Ca class=\"anchor\" href=\"#4-量化粒度-逐层-通道-分组\" aria-label=\"跳转到标题：4. 量化粒度（逐层\u002F通道\u002F分组）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. 量化粒度（逐层\u002F通道\u002F分组）\u003C\u002Fh4>\u003Col>\n\u003Cli>张量级：整张张量共用一组 scale，实现最简单；\u003C\u002Fli>\n\u003Cli>通道级：每个通道独立缩放，精度提升；\u003C\u002Fli>\n\u003Cli>组级（MXFP4 等方案）：多个连续元素共享一个缩放因子，兼顾显存与精度。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw10.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=jLusNbnv%2FF89L5UayCfFibXMG5g%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"int-8-量化完整执行流程\">\u003Ca class=\"anchor\" href=\"#int-8-量化完整执行流程\" aria-label=\"跳转到标题：INT 8 量化完整执行流程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>INT 8 量化完整执行流程\u003C\u002Fh3>\u003Col>\n\u003Cli>推理前：FP 32 权重经标定计算 scale \u002F zero，量化存储为 INT 8；\u003C\u002Fli>\n\u003Cli>推理前向：根据硬件和算子实现，可采用 INT8 输入直接进行整数计算，或者在计算前转换为低精度浮点执行；\u003C\u002Fli>\n\u003Cli>进阶优化：全程 INT 8 向量化访存与计算，减少 HBM 读写次数；\u003C\u002Fli>\n\u003Cli>性能收益：同等硬件下吞吐相比 FP 16 可提升近一倍，代价为有限精度损耗。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw11.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=g%2FjYtWUxqZQlY8kEY8vpobjr4hA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw12.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=vQ47P21BsTy2KBqRHWI8h8VqQ%2B0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"量化取舍原则\">\u003Ca class=\"anchor\" href=\"#量化取舍原则\" aria-label=\"跳转到标题：量化取舍原则\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>量化取舍原则\u003C\u002Fh3>\u003Cul>\n\u003Cli>量化依靠牺牲部分数值精度换取显存、算力收益；\u003C\u002Fli>\n\u003Cli>文本、图像生成任务可适度采用，金融、仿真等强精度场景不建议无限制低位宽量化。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw13.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=9m%2FzTLFWpvr0BddLxnl0lXpQXTE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"性能优化-ptx\">\u003Ca class=\"anchor\" href=\"#性能优化-ptx\" aria-label=\"跳转到标题：性能优化——PTX\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>性能优化——PTX\u003C\u002Fh2>\u003Ch3 id=\"ptx-核心价值\">\u003Ca class=\"anchor\" href=\"#ptx-核心价值\" aria-label=\"跳转到标题：PTX 核心价值\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>PTX 核心价值\u003C\u002Fh3>\u003Cp>介于 C++ 与硬件指令之间的虚拟汇编层：\u003C\u002Fp>\n\u003Col>\n\u003Cli>提供上层代码未暴露的细粒度硬件操作；\u003C\u002Fli>\n\u003Cli>规避编译器保守优化，手动调度 MMA、异步拷贝等专用指令；\u003C\u002Fli>\n\u003Cli>跨 GPU 架构向前兼容，提高代码迁移灵活性。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"典型-ptx-落地项目案例\">\u003Ca class=\"anchor\" href=\"#典型-ptx-落地项目案例\" aria-label=\"跳转到标题：典型 PTX 落地项目案例\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>典型 PTX 落地项目案例\u003C\u002Fh3>\u003Col>\n\u003Cli>DeepSeek MoE 通信库：自定义分组门控 PTX 汇编优化通信开销；\u003C\u002Fli>\n\u003Cli>DeepGEMM FP 8 矩阵乘：利用 PTX 中的矩阵加载\u002F存储指令优化 Tensor Core 数据搬运；\u003C\u002Fli>\n\u003Cli>Flash LLM：Load\u002FStore 专用内存指令优化长上下文访存。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw14.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2FhFdV6wizQUZPeTUjzzG5dOnLug%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw15.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Q9DRHSVzssYYlZt7404sVUFkMwM%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw16.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hiqBPj%2F72R7Is9aKusMKt3WQ8b0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw17.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FMMDLwtD4kiRV3mbKEOj%2Fc0hn7E%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw18.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=p4ow%2BWgNNxNUKJVwFiyKb3Qtl9M%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw19.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sJPlOYS6pNJi%2BC3O%2FJLcMmBDKfE%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"runtime-vs-driver-api\">\u003Ca class=\"anchor\" href=\"#runtime-vs-driver-api\" aria-label=\"跳转到标题：Runtime vs Driver API\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Runtime vs Driver API\u003C\u002Fh3>\u003Cul>\n\u003Cli>Runtime API：高层封装，开发效率高，通用场景首选；\u003C\u002Fli>\n\u003Cli>Driver API：底层管控，支持动态加载、硬件精细调优，算子极致优化场景使用。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"微指令调优\">\u003Ca class=\"anchor\" href=\"#微指令调优\" aria-label=\"跳转到标题：微指令调优\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>微指令调优\u003C\u002Fh2>\u003Ch3 id=\"循环优化指令\">\u003Ca class=\"anchor\" href=\"#循环优化指令\" aria-label=\"跳转到标题：循环优化指令\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>循环优化指令\u003C\u002Fh3>\u003Cp>\u003Ccode>#pragma unroll\u003C\u002Fcode> 循环展开：消除循环分支开销、提升 ILP 指令并行；\u003Cbr>注意过度展开会造成代码膨胀、寄存器溢出，需控制展开倍数。\u003C\u002Fp>\n\u003Ch3 id=\"断言与约束修饰符\">\u003Ca class=\"anchor\" href=\"#断言与约束修饰符\" aria-label=\"跳转到标题：断言与约束修饰符\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>断言与约束修饰符\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Ccode>__builtin_assume(cond)\u003C\u002Fcode>：向编译器提供条件约束信息，帮助优化代码路径；\u003C\u002Fli>\n\u003Cli>\u003Ccode>__align__\u003C\u002Fcode>：指定数据对齐要求，提高满足硬件访存对齐条件的可能性。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch3 id=\"缓存控制编译选项\">\u003Ca class=\"anchor\" href=\"#缓存控制编译选项\" aria-label=\"跳转到标题：缓存控制编译选项\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>缓存控制编译选项\u003C\u002Fh3>\u003Col>\n\u003Cli>CA（默认）：完整启用 L1 缓存；\u003C\u002Fli>\n\u003Cli>CG：绕过 L1，全部资源分配给共享内存；\u003C\u002Fli>\n\u003Cli>CS：流式内存标记，缓存数据使用后立即丢弃，减少缓存污染。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw20.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=VAFYnhBnQ9yuEdERscR5T%2FctiyQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"浮点与计算编译开关\">\u003Ca class=\"anchor\" href=\"#浮点与计算编译开关\" aria-label=\"跳转到标题：浮点与计算编译开关\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>浮点与计算编译开关\u003C\u002Fh3>\u003Col>\n\u003Cli>\u003Ccode>--use_fast_math\u003C\u002Fcode>：近似三角函数，提速小幅损失精度；\u003C\u002Fli>\n\u003Cli>\u003Ccode>--ftz=true\u003C\u002Fcode>：刷新非规格浮点数为 0，简化计算；\u003Cbr>3 寄存器上限限制：防止寄存器溢出至本地内存拖慢速度。\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2 id=\"warp-线程发散问题与优化\">\u003Ca class=\"anchor\" href=\"#warp-线程发散问题与优化\" aria-label=\"跳转到标题：Warp 线程发散问题与优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Warp 线程发散问题与优化\u003C\u002Fh2>\u003Ch3 id=\"发散成因\">\u003Ca class=\"anchor\" href=\"#发散成因\" aria-label=\"跳转到标题：发散成因\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>发散成因\u003C\u002Fh3>\u003Cp>同一 Warp（32 线程）内执行不同分支路径，GPU 串行执行所有分支，硬件利用率暴跌，NCU 可观测长 Stall 指标。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw21.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=j3XS39uDgUg1QkUbkHdTROGUYk8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw22.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=x9IgbTB5E%2BBn%2BjkaCkOb4GKBrXQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"主流规避方案\">\u003Ca class=\"anchor\" href=\"#主流规避方案\" aria-label=\"跳转到标题：主流规避方案\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>主流规避方案\u003C\u002Fh3>\u003Col>\n\u003Cli>Mask 掩码计算：所有线程完整执行计算，通过掩码控制写入，消除分支跳转；\u003C\u002Fli>\n\u003Cli>算法重构：拆分条件逻辑，对齐线程执行路径；\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw23.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=D5T%2BZtJXBvXagJQ1IrwCqzeDNj8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw24.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=SOwnFKQaCM81JsOvOazRcysAWC8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw25.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4bPkd53B5rTx9khvQI%2FzMJkRE4Y%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw26.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=qWtUTL0Y7u2fBmVtkCUswqvufCQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw27.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=CuK2hIFNrASJur3PTPgOdxguaAk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"指标观测\">\u003Ca class=\"anchor\" href=\"#指标观测\" aria-label=\"跳转到标题：指标观测\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>指标观测\u003C\u002Fh3>\u003Cp>NCU 查看 \u003Ccode>average predicate thread\u003C\u002Fcode> 平均激活线程数，数值越低发散越严重。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw28.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hK46sgOKmMH6CzgTYeiyukKxNw0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"并行平衡与数据排布调优\">\u003Ca class=\"anchor\" href=\"#并行平衡与数据排布调优\" aria-label=\"跳转到标题：并行平衡与数据排布调优\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>并行平衡与数据排布调优\u003C\u002Fh2>\u003Ch3 id=\"ilp-指令级并行\">\u003Ca class=\"anchor\" href=\"#ilp-指令级并行\" aria-label=\"跳转到标题：ILP 指令级并行\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>ILP 指令级并行\u003C\u002Fh3>\u003Cp>无依赖指令重叠执行，利用访存等待周期同步计算；手写 PTX 可手动调度指令顺序放大 ILP 收益。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw29.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=uzgLV524%2FzOyJm%2FvH%2BUfRedMV54%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"tlp-线程级并行\">\u003Ca class=\"anchor\" href=\"#tlp-线程级并行\" aria-label=\"跳转到标题：TLP 线程级并行\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>TLP 线程级并行\u003C\u002Fh3>\u003Cp>多线程分摊计算负载；ILP 与 TLP 需要平衡，单一维度拉满会占用过多寄存器\u002F共享内存。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw30.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=HIG3uRvxmOxRwaqg6bBhuJUi9j4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw31.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=A9DuxBGX6WSU%2FD2YDv4qdEO1qIs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw32.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=4tEccuHOKzEyEE7LBvQQLXKbAf4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"aos-soa-数据结构选择\">\u003Ca class=\"anchor\" href=\"#aos-soa-数据结构选择\" aria-label=\"跳转到标题：AOS \u002F SOA 数据结构选择\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>AOS \u002F SOA 数据结构选择\u003C\u002Fh3>\u003Cul>\n\u003Cli>AOS：结构体数组，随机访存友好，批量连续读取效率差；\u003C\u002Fli>\n\u003Cli>SOA：数组结构体，同字段连续排布，向量化加载、矩阵乘场景性能显著优于 AOS。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw33.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=S%2BHshDzz6MTEqTm1pmei9sIkTPk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"block-size-调优\">\u003Ca class=\"anchor\" href=\"#block-size-调优\" aria-label=\"跳转到标题：Block Size 调优\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>Block Size 调优\u003C\u002Fh3>\u003Cp>默认 256 非最优，不同算子、硬件架构适配最优 Block 尺寸，需结合寄存器占用、共享内存上限测试。\u003C\u002Fp>\n\u003Ch2 id=\"cuda-官方库\">\u003Ca class=\"anchor\" href=\"#cuda-官方库\" aria-label=\"跳转到标题：CUDA 官方库\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA 官方库\u003C\u002Fh2>\u003Ch4 id=\"1-cudnn\">\u003Ca class=\"anchor\" href=\"#1-cudnn\" aria-label=\"跳转到标题：1. cuDNN\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>1. cuDNN\u003C\u002Fh4>\u003Cp>深度神经网络专用库，卷积、归一化、激活高度优化，原生支持 Tensor Core 混合精度、算子融合。\u003C\u002Fp>\n\u003Ch4 id=\"2-cublas\">\u003Ca class=\"anchor\" href=\"#2-cublas\" aria-label=\"跳转到标题：2. cuBLAS\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>2. cuBLAS\u003C\u002Fh4>\u003Cp>标准线性代数库，向量\u002F矩阵乘全精度实现。\u003C\u002Fp>\n\u003Ch4 id=\"3-cub\">\u003Ca class=\"anchor\" href=\"#3-cub\" aria-label=\"跳转到标题：3. CUB\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>3. CUB\u003C\u002Fh4>\u003Cp>CUDA 高性能并行原语库，提供 Block、Warp、Device 级别的扫描、归约、排序等基础操作。\u003C\u002Fp>\n\u003Ch4 id=\"4-cutlass\">\u003Ca class=\"anchor\" href=\"#4-cutlass\" aria-label=\"跳转到标题：4. CUTLASS\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>4. CUTLASS\u003C\u002Fh4>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw34.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=sg7njRdIe1%2BZWWQnikuUuSLz1Ko%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch4 id=\"5-cusparse\">\u003Ca class=\"anchor\" href=\"#5-cusparse\" aria-label=\"跳转到标题：5.cuSPARSE\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>5.cuSPARSE\u003C\u002Fh4>\u003Cp>稀疏矩阵专用运算库，支持多种稀疏存储格式与分解求解。\u003C\u002Fp>\n\u003Ch2 id=\"cuda-全系列课程完整复盘\">\u003Ca class=\"anchor\" href=\"#cuda-全系列课程完整复盘\" aria-label=\"跳转到标题：CUDA 全系列课程完整复盘\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CUDA 全系列课程完整复盘\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_98_zw35.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=MFvf09VwvzDDYlZR471eKB9iUrs%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>\u003Cp>📺\u003Cstrong>完整课程内容，请观看直播或查看课程回放：\u003C\u002Fstrong>\u003C\u002Fp>\n\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\u002FKnowledge_Base\u002Fqun_1072.jpg?Expires=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=OqqoZKUCZDwCVp9tER7EkbY9quU%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=1792750031&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=G1NJskF7AaercUrU0ISrAy2%2F%2F6c%3D\" alt=\"\">\u003C\u002Fp>\n",[128,130,132,134,136,138,142,145,148,151,154,156,159,162,165,168,170,172,174,176,178,181,183,185,187,189,192,195,198,201,204,207,210,213,216,219,222,224],{"id":129,"text":129,"level":110},"低精度浮点数据类型完整解析",{"id":131,"text":131,"level":9},"主流低精度类型特性对比",{"id":133,"text":133,"level":110},"深度学习量化完整技术体系",{"id":135,"text":135,"level":9},"量化基础定义",{"id":137,"text":137,"level":9},"量化四大分类维度",{"id":139,"text":140,"level":141},"1-线性-vs-非线性量化","1. 线性 vs 非线性量化",4,{"id":143,"text":144,"level":141},"2-对称-vs-非对称量化","2. 对称 vs 非对称量化",{"id":146,"text":147,"level":141},"3-静态-vs-动态量化","3. 静态 vs 动态量化",{"id":149,"text":150,"level":141},"4-量化粒度-逐层-通道-分组","4. 量化粒度（逐层\u002F通道\u002F分组）",{"id":152,"text":153,"level":9},"int-8-量化完整执行流程","INT 8 量化完整执行流程",{"id":155,"text":155,"level":9},"量化取舍原则",{"id":157,"text":158,"level":110},"性能优化-ptx","性能优化——PTX",{"id":160,"text":161,"level":9},"ptx-核心价值","PTX 核心价值",{"id":163,"text":164,"level":9},"典型-ptx-落地项目案例","典型 PTX 落地项目案例",{"id":166,"text":167,"level":9},"runtime-vs-driver-api","Runtime vs Driver API",{"id":169,"text":169,"level":110},"微指令调优",{"id":171,"text":171,"level":9},"循环优化指令",{"id":173,"text":173,"level":9},"断言与约束修饰符",{"id":175,"text":175,"level":9},"缓存控制编译选项",{"id":177,"text":177,"level":9},"浮点与计算编译开关",{"id":179,"text":180,"level":110},"warp-线程发散问题与优化","Warp 线程发散问题与优化",{"id":182,"text":182,"level":9},"发散成因",{"id":184,"text":184,"level":9},"主流规避方案",{"id":186,"text":186,"level":9},"指标观测",{"id":188,"text":188,"level":110},"并行平衡与数据排布调优",{"id":190,"text":191,"level":9},"ilp-指令级并行","ILP 指令级并行",{"id":193,"text":194,"level":9},"tlp-线程级并行","TLP 线程级并行",{"id":196,"text":197,"level":9},"aos-soa-数据结构选择","AOS \u002F SOA 数据结构选择",{"id":199,"text":200,"level":9},"block-size-调优","Block Size 调优",{"id":202,"text":203,"level":110},"cuda-官方库","CUDA 官方库",{"id":205,"text":206,"level":141},"1-cudnn","1. cuDNN",{"id":208,"text":209,"level":141},"2-cublas","2. cuBLAS",{"id":211,"text":212,"level":141},"3-cub","3. CUB",{"id":214,"text":215,"level":141},"4-cutlass","4. CUTLASS",{"id":217,"text":218,"level":141},"5-cusparse","5.cuSPARSE",{"id":220,"text":221,"level":110},"cuda-全系列课程完整复盘","CUDA 全系列课程完整复盘",{"id":223,"text":223,"level":110},"加入训练营",{"id":225,"text":225,"level":110},"关注与交流"]