[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"site-home-live-entry-setting":3,"site-header-contests":8,"site-header-camps":49,"study-material-detail-study-87":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_87_fm.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mKNjR07UuWksajmtVUaPqUC7Sa4%3D","寒武纪 NeuWare,BANG,MLU,CNCL,Triton 编译器","全面讲解寒武纪 NeuWare 五层软件栈与 BANG 异构平台，解析 CNToolkit 工具链、CNCL 通信库、Triton 适配方案，覆盖 MLU 硬件架构、算子开发与多框架适配技术。","2026-10-22 20:38:11",2,"2026 夏季 InfiniTensor 训练营 CUDA 编程 6：《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 1》","![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mai5%2FpIwjvVzeIy9IpNxGEOj2x4%3D)\r\n\r\n\u003Cspan style=\"font-size:12px;color:#555;\">7 月 31 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第六课《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 1》开讲。\r\n前置课程《内存模型与规约优化》重点讲解并行规约中的原子操作、GPU 内存层级架构；《分块与不规则访存》解析 CUDA GEMM 矩阵乘法极致优化与 Bank Conflict 解决方案，涵盖共享内存机制、动态并行等核心技术。\u003C\u002Fspan>\r\n\r\n本节课由寒武纪团队带来 NeuWare 软件栈的层次结构及各层协作机制，以及 BANG 异构计算平台配套的工具链（编译、调试、性能分析等）的深度解析。为后续 BANG C++ 算子开发、并行编程模型及性能优化课程打下理论基础。\r\n\r\n## NeuWare 软件栈全貌解析\r\n### 软件栈分层架构\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Dyt%2FRelRH61PMl9uhhqrtCX18K8%3D)\r\n\r\n1.**应用层**：端云一体支持（互联网\u002F金融\u002F交通等）\r\n\r\n2.**框架层**：\r\n- 开源框架支持：PyTorch、TensorFlow 等\r\n- 寒武纪扩展组件\r\n\r\n3.**计算库层**：\r\n- NN 算子库（CNL）、CV 算子库、通信库（CNCL）。\r\n- 配套工具：拓扑查看工具、图形化调优工具（Perf）。\r\n\r\n4.**BANG 异构计算平台**：\r\n- 核心工具包 CNToolkit：编译器、运行时、编解码、性能调试工具。\r\n- 算子库与通信库均基于此工具链开发。\r\n\r\n5.**驱动层**：\r\n- 多平台驱动（端云一体）、硬件抽象整合。\r\n- 云端部署工具：质量检测、验证套件。\r\n\r\n### 开发者定位指南\r\n- **算法工程师**：聚焦开源 AI 框架层。\r\n- **框架定制开发者**：需掌握计算库与工具使用。\r\n- **算子\u002F底层开发者**：需精通 BANG 编程语言、调优工具及汇编级优化。\r\n\r\n\r\n## 多视角下的软件栈技术细节\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=S0D%2FfqzXFfFumVi3NHN3c4VGFB8%3D)\r\n\r\n### 硬件与基础软件协同\r\n1.**硬件技术**：微架构指令集、SoC 设计、板卡\u002F集群\u002F超节点技术。\r\n\r\n2.**基础软件**：自研编译器、汇编器、二进制格式、链接器、BANGC\u002FTriton 语言。\r\n\r\n3.**关键抽象层**：\r\n- 驱动层：屏蔽多款硬件差异。\r\n- BANG 平台层：提供异构并行编程模型（类比 CUDA）。\r\n\r\n### 开发者视角的软件栈\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3zo9t3%2Fs6gODifI1Ag8FD2sd924%3D)\r\n\r\n### 工具链视角\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8JmfZBpN0l5xFJ3i6F7oIyX%2B2Xw%3D)\r\n\r\n\r\n## SDK 演进与编译范式转移\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=apZHpmKeEIlvEIPcgbxz3uvtEak%3D)\r\n\r\n\r\n\r\n\r\n1.**历史演进**：从定制化库 → SDK 1.0 基础软件 → 多层 IR 生态。\r\n\r\n2.**多层 IR 体系**：\r\n- 框架层 IR（HLO\u002FTriton IR）→ MLU VM IR → MELISSA 汇编。\r\n- 各层各司其职，兼顾生态泛化性与性能。\r\n\r\n3.**编译范式**：\r\n- MagicMind 闭源引擎 → MLIR 开源生态对接。\r\n- Triton\u002FLinalg → MRO → MLVM → PTX 类汇编（与 CNCC 共用后端）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6o1jR5VOssG88RmI1Vv%2FXon0GM0%3D)\r\n\r\n\r\n## AI 框架与开源生态\r\n### 主流框架适配\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NgwN30a4I3lmqJT8B9CoAmkJmPk%3D)\r\n\r\n1.**PyTorch**：\r\n- 非侵入式接入（Private Key 注入），兼容动态图\u002F分布式\u002F Compile。\r\n\r\n2.**TensorFlow**：\r\n- Cambricon-TF 包，设备名替换即可。\r\n- 支持 XLA 自动编译、MagicMind 部署转换。\r\n\r\n### 训练生态\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0SABDWDJjMs%2BanyYMFvJguB70qA%3D)\r\n\r\n1.**训练生态**：Megatron、DeepSpeed、FA v2、Lightning 等原生支持。\r\n- 透传机制：高层框架 → Transformer Engine\u002FApex → CNL\u002FCNCL\u002FExtra 库。\r\n\r\n2.**强化学习\u002F vLLM**：\r\n- 热点算子集成至 `torch_mlu_ops`。\r\n- vLLM 适配：支持多机多卡、MRGraph（对标 CUDA Graph）、KV Cache INT8。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l82hZNQ%2BkUDJnKPWjy9T5eaHvA0%3D)\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RiTkt%2FcA86Lb9SUfn%2BcF3ctvVDs%3D)\r\n\r\n3.**Triton 生态**：\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FuKGdHwV3rBHjLdCJ2JVhHbmIVI%3D)\r\n\r\n- 原生支持 TL 原语，性能达 BANG C 极致优化的 80%。\r\n- 开发周期从 10 天缩短至 1-2 天（配合 Skill 工具）。\r\n\r\n### 专用加速库\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PIONtSI8HGSCEmXRJQQgO2pjPnQ%3D)\r\n\r\n- **torch_mlu_ops**：无缝对接 PyTorch Compile，集成 FA\u002F量化算子。\r\n- **CNTrainKit**：训练前性能预估、并行策略排布、静态内存分析。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1EtO7vYPF0wM%2F6SJ9ujDu5cfJFo%3D)\r\n\r\n\r\n\r\n## 加速库层核心技术\r\n### 计算加速范式\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QxcSl7Z%2BM5UNAVmgtvO2ofuYPQU%3D)\r\n\r\n1.**乒乓流水**：双片上 RAM + 计算\u002FIO 流并行，Load-Compute-Store 全重叠。\r\n\r\n2.**工具链**：CNPerf 抓 PMU 数据 + GUI 分析热点。\r\n\r\n### 通信加速范式\r\n1.**CNCL**：对标 NCCL，支持 AllGather\u002FReduce 等原语，基于 BANG C + CNDRV 开发。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BOWAZ8DQpsPuCDwaAa31fFYxZsA%3D)\r\n\r\n2.**CNSHMEM**：对标 nvshmem，PGAS 编程模型，支持 MoE 稀疏通信与通算融合。\r\n- 与 PyTorch Symmetric Memory + Triton JIT 联动。\r\n\r\n### CNTLASS 模板库\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2AhM6GaYaPDXJyNW3TL9V5Zx0FU%3D)\r\n\r\n- **高层抽象**：Tile\u002FCNTE 编程模型，C++ 语法封装 Tensor\u002FLayout。\r\n- **开发效率**：Attention 算子开发从 1 个月缩短至 1 周。\r\n- **性能表现**：BF16 GEMM 逼近硬件极限，支持 Autotuning。\r\n\r\n\r\n## BANG 异构计算平台深度解析\r\n### 核心工具链（CNToolkit）\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bhtkw11Xn69dmcQccGH0ZrH2DvQ%3D)\r\n\r\n- **CNCC**：BANG C 编译器，支持异构混合编码，Host\u002FDevice 自动拆分。\r\n- **CNPerf**：Kernel 级性能分析（执行时间\u002F指令数\u002FCache Miss）。\r\n- **Sanitizer**：内存越界\u002F未定义行为检测（MemCheck\u002FRaceCheck）。\r\n- **CNGDB\u002FStudio**：调试器与 VS Code 插件。\r\n\r\n###  硬件架构与存储模型\r\n1.**Cluster 结构**：IPU\u002FMPU + NRAM\u002FWRAM\u002FSRAM + 功能单元（向量\u002F标量\u002F张量）。\r\n\r\n2.**存储层级**：\r\n  - 片上：NRAM\u002FWRAM\u002FSRAM（异步搬移，增强数据局部性）。\r\n  - 片外：Global DRAM + Local DRAM（线程栈）+ L2 Cache。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hnnlTM6JMTjstM59I77YOo2pX18%3D)\r\n\r\n### 执行模型与并行编程\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=U%2BKTI%2Ftpo671SzbU2ygZ0yYwGVQ%3D)\r\n\r\n1.**Host-Device 异步**：独立执行，显式 Sync 同步。\r\n\r\n2.**Block\u002FUnit 映射**：Kernel 按 Unit 粒度占用 Cluster，支持排队复用。\r\n\r\n3.**内建变量**：TaskID\u002FTaskDim\u002FCallID 索引并行任务。\r\n\r\n### 加速范式与编译优化\r\n1.**图执行**：基于阿姆达尔\u002F古斯塔夫森定律，压缩串行部分，最大化并行。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bRLqlyUrcX2bEHQZeE7SdlKGtZU%3D)\r\n\r\n2.**算子融合**：\r\n- 编译时融合：Inductor\u002FXLA 自动生成 Triton Kernel。\r\n- 手写融合：BANG C 大 Kernel 减少片外访存。\r\n\r\n## 驱动层关键技术\r\n### 核心功能\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ASzxTPuzhSfDcYcU6v4RYRz7xmo%3D)\r\n\r\n1.**虚拟化**：单卡切分多虚拟卡（物理隔离）。\r\n\r\n2.**软件隔离**：Cluster Bitmap 环境变量控制（类似 CUDA MPS）。\r\n\r\n![](https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw23.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mH6rDYlqhMVPgXXu1%2F5i%2BMzfITg%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\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-86","study-86","2026 夏季 InfiniTensor 训练营 Triton & 九齿 4：《九齿二重：渐悟》",22,"\u002Fstudy-material\u002Fstudy-87","study-87","markdown",{"publishTime":7,"createTime":109,"canonicalPath":122,"id":123,"title":124},"\u002Fstudy-material\u002Fstudy-88","study-88","2026 夏季 InfiniTensor 训练营推理引擎 4：《大模型计算优化与分布式推理》",{"htmlContent":126,"headings":127},"\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw01.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mai5%2FpIwjvVzeIy9IpNxGEOj2x4%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cspan style=\"font-size:12px;color:#555;\">7 月 31 日，2026 夏季 InfiniTensor 训练营 CUDA 编程方向第六课《寒武纪软件栈 NeuWare 和 BANG 异构计算平台 1》开讲。\u003Cbr>前置课程《内存模型与规约优化》重点讲解并行规约中的原子操作、GPU 内存层级架构；《分块与不规则访存》解析 CUDA GEMM 矩阵乘法极致优化与 Bank Conflict 解决方案，涵盖共享内存机制、动态并行等核心技术。\u003C\u002Fspan>\u003C\u002Fp>\n\u003Cp>本节课由寒武纪团队带来 NeuWare 软件栈的层次结构及各层协作机制，以及 BANG 异构计算平台配套的工具链（编译、调试、性能分析等）的深度解析。为后续 BANG C++ 算子开发、并行编程模型及性能优化课程打下理论基础。\u003C\u002Fp>\n\u003Ch2 id=\"neuware-软件栈全貌解析\">\u003Ca class=\"anchor\" href=\"#neuware-软件栈全貌解析\" aria-label=\"跳转到标题：NeuWare 软件栈全貌解析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>NeuWare 软件栈全貌解析\u003C\u002Fh2>\u003Ch3 id=\"软件栈分层架构\">\u003Ca class=\"anchor\" href=\"#软件栈分层架构\" aria-label=\"跳转到标题：软件栈分层架构\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>软件栈分层架构\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw02.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=Dyt%2FRelRH61PMl9uhhqrtCX18K8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>应用层\u003C\u002Fstrong>：端云一体支持（互联网\u002F金融\u002F交通等）\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>框架层\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>开源框架支持：PyTorch、TensorFlow 等\u003C\u002Fli>\n\u003Cli>寒武纪扩展组件\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>3.\u003Cstrong>计算库层\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>NN 算子库（CNL）、CV 算子库、通信库（CNCL）。\u003C\u002Fli>\n\u003Cli>配套工具：拓扑查看工具、图形化调优工具（Perf）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>4.\u003Cstrong>BANG 异构计算平台\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>核心工具包 CNToolkit：编译器、运行时、编解码、性能调试工具。\u003C\u002Fli>\n\u003Cli>算子库与通信库均基于此工具链开发。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>5.\u003Cstrong>驱动层\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>多平台驱动（端云一体）、硬件抽象整合。\u003C\u002Fli>\n\u003Cli>云端部署工具：质量检测、验证套件。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"开发者定位指南\">\u003Ca class=\"anchor\" href=\"#开发者定位指南\" aria-label=\"跳转到标题：开发者定位指南\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>开发者定位指南\u003C\u002Fh3>\u003Cul>\n\u003Cli>\u003Cstrong>算法工程师\u003C\u002Fstrong>：聚焦开源 AI 框架层。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>框架定制开发者\u003C\u002Fstrong>：需掌握计算库与工具使用。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>算子\u002F底层开发者\u003C\u002Fstrong>：需精通 BANG 编程语言、调优工具及汇编级优化。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"多视角下的软件栈技术细节\">\u003Ca class=\"anchor\" href=\"#多视角下的软件栈技术细节\" aria-label=\"跳转到标题：多视角下的软件栈技术细节\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>多视角下的软件栈技术细节\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw03.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=S0D%2FfqzXFfFumVi3NHN3c4VGFB8%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"硬件与基础软件协同\">\u003Ca class=\"anchor\" href=\"#硬件与基础软件协同\" aria-label=\"跳转到标题：硬件与基础软件协同\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>硬件与基础软件协同\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>硬件技术\u003C\u002Fstrong>：微架构指令集、SoC 设计、板卡\u002F集群\u002F超节点技术。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>基础软件\u003C\u002Fstrong>：自研编译器、汇编器、二进制格式、链接器、BANGC\u002FTriton 语言。\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>关键抽象层\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>驱动层：屏蔽多款硬件差异。\u003C\u002Fli>\n\u003Cli>BANG 平台层：提供异构并行编程模型（类比 CUDA）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"开发者视角的软件栈\">\u003Ca class=\"anchor\" href=\"#开发者视角的软件栈\" aria-label=\"跳转到标题：开发者视角的软件栈\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>开发者视角的软件栈\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw04.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=3zo9t3%2Fs6gODifI1Ag8FD2sd924%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"工具链视角\">\u003Ca class=\"anchor\" href=\"#工具链视角\" aria-label=\"跳转到标题：工具链视角\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>工具链视角\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw05.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=8JmfZBpN0l5xFJ3i6F7oIyX%2B2Xw%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"sdk-演进与编译范式转移\">\u003Ca class=\"anchor\" href=\"#sdk-演进与编译范式转移\" aria-label=\"跳转到标题：SDK 演进与编译范式转移\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>SDK 演进与编译范式转移\u003C\u002Fh2>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw06.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=apZHpmKeEIlvEIPcgbxz3uvtEak%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>历史演进\u003C\u002Fstrong>：从定制化库 → SDK 1.0 基础软件 → 多层 IR 生态。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>多层 IR 体系\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>框架层 IR（HLO\u002FTriton IR）→ MLU VM IR → MELISSA 汇编。\u003C\u002Fli>\n\u003Cli>各层各司其职，兼顾生态泛化性与性能。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>3.\u003Cstrong>编译范式\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>MagicMind 闭源引擎 → MLIR 开源生态对接。\u003C\u002Fli>\n\u003Cli>Triton\u002FLinalg → MRO → MLVM → PTX 类汇编（与 CNCC 共用后端）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw07.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=6o1jR5VOssG88RmI1Vv%2FXon0GM0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"ai-框架与开源生态\">\u003Ca class=\"anchor\" href=\"#ai-框架与开源生态\" aria-label=\"跳转到标题：AI 框架与开源生态\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>AI 框架与开源生态\u003C\u002Fh2>\u003Ch3 id=\"主流框架适配\">\u003Ca class=\"anchor\" href=\"#主流框架适配\" aria-label=\"跳转到标题：主流框架适配\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>主流框架适配\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw08.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=NgwN30a4I3lmqJT8B9CoAmkJmPk%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>PyTorch\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>非侵入式接入（Private Key 注入），兼容动态图\u002F分布式\u002F Compile。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>TensorFlow\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>Cambricon-TF 包，设备名替换即可。\u003C\u002Fli>\n\u003Cli>支持 XLA 自动编译、MagicMind 部署转换。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"训练生态\">\u003Ca class=\"anchor\" href=\"#训练生态\" aria-label=\"跳转到标题：训练生态\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>训练生态\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw09.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=0SABDWDJjMs%2BanyYMFvJguB70qA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>训练生态\u003C\u002Fstrong>：Megatron、DeepSpeed、FA v2、Lightning 等原生支持。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>透传机制：高层框架 → Transformer Engine\u002FApex → CNL\u002FCNCL\u002FExtra 库。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>2.\u003Cstrong>强化学习\u002F vLLM\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>热点算子集成至 \u003Ccode>torch_mlu_ops\u003C\u002Fcode>。\u003C\u002Fli>\n\u003Cli>vLLM 适配：支持多机多卡、MRGraph（对标 CUDA Graph）、KV Cache INT8。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw10.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=l82hZNQ%2BkUDJnKPWjy9T5eaHvA0%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw11.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=RiTkt%2FcA86Lb9SUfn%2BcF3ctvVDs%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>Triton 生态\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw12.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=FuKGdHwV3rBHjLdCJ2JVhHbmIVI%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>原生支持 TL 原语，性能达 BANG C 极致优化的 80%。\u003C\u002Fli>\n\u003Cli>开发周期从 10 天缩短至 1-2 天（配合 Skill 工具）。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"专用加速库\">\u003Ca class=\"anchor\" href=\"#专用加速库\" aria-label=\"跳转到标题：专用加速库\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>专用加速库\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw13.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=PIONtSI8HGSCEmXRJQQgO2pjPnQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>torch_mlu_ops\u003C\u002Fstrong>：无缝对接 PyTorch Compile，集成 FA\u002F量化算子。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>CNTrainKit\u003C\u002Fstrong>：训练前性能预估、并行策略排布、静态内存分析。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw14.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=1EtO7vYPF0wM%2F6SJ9ujDu5cfJFo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch2 id=\"加速库层核心技术\">\u003Ca class=\"anchor\" href=\"#加速库层核心技术\" aria-label=\"跳转到标题：加速库层核心技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加速库层核心技术\u003C\u002Fh2>\u003Ch3 id=\"计算加速范式\">\u003Ca class=\"anchor\" href=\"#计算加速范式\" aria-label=\"跳转到标题：计算加速范式\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>计算加速范式\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw15.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=QxcSl7Z%2BM5UNAVmgtvO2ofuYPQU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>乒乓流水\u003C\u002Fstrong>：双片上 RAM + 计算\u002FIO 流并行，Load-Compute-Store 全重叠。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>工具链\u003C\u002Fstrong>：CNPerf 抓 PMU 数据 + GUI 分析热点。\u003C\u002Fp>\n\u003Ch3 id=\"通信加速范式\">\u003Ca class=\"anchor\" href=\"#通信加速范式\" aria-label=\"跳转到标题：通信加速范式\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>通信加速范式\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>CNCL\u003C\u002Fstrong>：对标 NCCL，支持 AllGather\u002FReduce 等原语，基于 BANG C + CNDRV 开发。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw16.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2BOWAZ8DQpsPuCDwaAa31fFYxZsA%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>CNSHMEM\u003C\u002Fstrong>：对标 nvshmem，PGAS 编程模型，支持 MoE 稀疏通信与通算融合。\u003C\u002Fp>\n\u003Cul>\n\u003Cli>与 PyTorch Symmetric Memory + Triton JIT 联动。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"cntlass-模板库\">\u003Ca class=\"anchor\" href=\"#cntlass-模板库\" aria-label=\"跳转到标题：CNTLASS 模板库\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>CNTLASS 模板库\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw17.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=2AhM6GaYaPDXJyNW3TL9V5Zx0FU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>高层抽象\u003C\u002Fstrong>：Tile\u002FCNTE 编程模型，C++ 语法封装 Tensor\u002FLayout。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>开发效率\u003C\u002Fstrong>：Attention 算子开发从 1 个月缩短至 1 周。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>性能表现\u003C\u002Fstrong>：BF16 GEMM 逼近硬件极限，支持 Autotuning。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"bang-异构计算平台深度解析\">\u003Ca class=\"anchor\" href=\"#bang-异构计算平台深度解析\" aria-label=\"跳转到标题：BANG 异构计算平台深度解析\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>BANG 异构计算平台深度解析\u003C\u002Fh2>\u003Ch3 id=\"核心工具链-cntoolkit\">\u003Ca class=\"anchor\" href=\"#核心工具链-cntoolkit\" aria-label=\"跳转到标题：核心工具链（CNToolkit）\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心工具链（CNToolkit）\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw18.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=%2Bhtkw11Xn69dmcQccGH0ZrH2DvQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>CNCC\u003C\u002Fstrong>：BANG C 编译器，支持异构混合编码，Host\u002FDevice 自动拆分。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>CNPerf\u003C\u002Fstrong>：Kernel 级性能分析（执行时间\u002F指令数\u002FCache Miss）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Sanitizer\u003C\u002Fstrong>：内存越界\u002F未定义行为检测（MemCheck\u002FRaceCheck）。\u003C\u002Fli>\n\u003Cli>\u003Cstrong>CNGDB\u002FStudio\u003C\u002Fstrong>：调试器与 VS Code 插件。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch3 id=\"硬件架构与存储模型\">\u003Ca class=\"anchor\" href=\"#硬件架构与存储模型\" aria-label=\"跳转到标题：硬件架构与存储模型\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>硬件架构与存储模型\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>Cluster 结构\u003C\u002Fstrong>：IPU\u002FMPU + NRAM\u002FWRAM\u002FSRAM + 功能单元（向量\u002F标量\u002F张量）。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>存储层级\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>片上：NRAM\u002FWRAM\u002FSRAM（异步搬移，增强数据局部性）。\u003C\u002Fli>\n\u003Cli>片外：Global DRAM + Local DRAM（线程栈）+ L2 Cache。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw19.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=hnnlTM6JMTjstM59I77YOo2pX18%3D\" alt=\"\">\u003C\u002Fp>\n\u003Ch3 id=\"执行模型与并行编程\">\u003Ca class=\"anchor\" href=\"#执行模型与并行编程\" aria-label=\"跳转到标题：执行模型与并行编程\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>执行模型与并行编程\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw20.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=U%2BKTI%2Ftpo671SzbU2ygZ0yYwGVQ%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>Host-Device 异步\u003C\u002Fstrong>：独立执行，显式 Sync 同步。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>Block\u002FUnit 映射\u003C\u002Fstrong>：Kernel 按 Unit 粒度占用 Cluster，支持排队复用。\u003C\u002Fp>\n\u003Cp>3.\u003Cstrong>内建变量\u003C\u002Fstrong>：TaskID\u002FTaskDim\u002FCallID 索引并行任务。\u003C\u002Fp>\n\u003Ch3 id=\"加速范式与编译优化\">\u003Ca class=\"anchor\" href=\"#加速范式与编译优化\" aria-label=\"跳转到标题：加速范式与编译优化\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>加速范式与编译优化\u003C\u002Fh3>\u003Cp>1.\u003Cstrong>图执行\u003C\u002Fstrong>：基于阿姆达尔\u002F古斯塔夫森定律，压缩串行部分，最大化并行。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw21.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=bRLqlyUrcX2bEHQZeE7SdlKGtZU%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>算子融合\u003C\u002Fstrong>：\u003C\u002Fp>\n\u003Cul>\n\u003Cli>编译时融合：Inductor\u002FXLA 自动生成 Triton Kernel。\u003C\u002Fli>\n\u003Cli>手写融合：BANG C 大 Kernel 减少片外访存。\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"驱动层关键技术\">\u003Ca class=\"anchor\" href=\"#驱动层关键技术\" aria-label=\"跳转到标题：驱动层关键技术\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>驱动层关键技术\u003C\u002Fh2>\u003Ch3 id=\"核心功能\">\u003Ca class=\"anchor\" href=\"#核心功能\" aria-label=\"跳转到标题：核心功能\">\u003Cspan class=\"octicon octicon-link\">\u003C\u002Fspan>\u003C\u002Fa>核心功能\u003C\u002Fh3>\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw22.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=ASzxTPuzhSfDcYcU6v4RYRz7xmo%3D\" alt=\"\">\u003C\u002Fp>\n\u003Cp>1.\u003Cstrong>虚拟化\u003C\u002Fstrong>：单卡切分多虚拟卡（物理隔离）。\u003C\u002Fp>\n\u003Cp>2.\u003Cstrong>软件隔离\u003C\u002Fstrong>：Cluster Bitmap 环境变量控制（类似 CUDA MPS）。\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"https:\u002F\u002Foss.infinitensor.com\u002Fprod\u002Fimages\u002FKnowledge_Base\u002FKB_87_zw23.jpg?Expires=1792752563&OSSAccessKeyId=LTAI5t9R6uzvTa69CT92cFvV&Signature=mH6rDYlqhMVPgXXu1%2F5i%2BMzfITg%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=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",[128,131,133,135,137,139,141,143,146,149,151,153,155,157,159,161,164,167,170,172,174,176,178,180,182,186],{"id":129,"text":130,"level":110},"neuware-软件栈全貌解析","NeuWare 软件栈全貌解析",{"id":132,"text":132,"level":9},"软件栈分层架构",{"id":134,"text":134,"level":9},"开发者定位指南",{"id":136,"text":136,"level":110},"多视角下的软件栈技术细节",{"id":138,"text":138,"level":9},"硬件与基础软件协同",{"id":140,"text":140,"level":9},"开发者视角的软件栈",{"id":142,"text":142,"level":9},"工具链视角",{"id":144,"text":145,"level":110},"sdk-演进与编译范式转移","SDK 演进与编译范式转移",{"id":147,"text":148,"level":110},"ai-框架与开源生态","AI 框架与开源生态",{"id":150,"text":150,"level":9},"主流框架适配",{"id":152,"text":152,"level":9},"训练生态",{"id":154,"text":154,"level":9},"专用加速库",{"id":156,"text":156,"level":110},"加速库层核心技术",{"id":158,"text":158,"level":9},"计算加速范式",{"id":160,"text":160,"level":9},"通信加速范式",{"id":162,"text":163,"level":9},"cntlass-模板库","CNTLASS 模板库",{"id":165,"text":166,"level":110},"bang-异构计算平台深度解析","BANG 异构计算平台深度解析",{"id":168,"text":169,"level":9},"核心工具链-cntoolkit","核心工具链（CNToolkit）",{"id":171,"text":171,"level":9},"硬件架构与存储模型",{"id":173,"text":173,"level":9},"执行模型与并行编程",{"id":175,"text":175,"level":9},"加速范式与编译优化",{"id":177,"text":177,"level":110},"驱动层关键技术",{"id":179,"text":179,"level":9},"核心功能",{"id":181,"text":181,"level":110},"加入训练营",{"id":183,"text":184,"level":185},"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":187,"text":187,"level":110},"关注与交流"]