{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-18T17:00:43.709Z","headline":"NVIDIA cuDNN Graph API 教程：融合、自动调优与 plan 复用","description":"一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。","url":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","mainEntityOfPage":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","datePublished":"2026-09-15T21:37:11.000Z","dateModified":"2026-09-15T21:37:11.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.marktechpost.com/2026/09/15/inside-nvidias-cudnn-graph-api-fusion-autotuning-and-plan-reuse-with-cudnn-frontend","https://aihot.news/items/cmu37dt9m0b2urosatw9cn0g8"],"canonicalUrl":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","dateCreated":"2026-09-15T21:37:11.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"marktechpost.com source article","url":"https://www.marktechpost.com/2026/09/15/inside-nvidias-cudnn-graph-api-fusion-autotuning-and-plan-reuse-with-cudnn-frontend","datePublished":"2026-09-15T21:37:11.000Z","provider":{"@type":"Organization","name":"marktechpost.com","url":"https://www.marktechpost.com/2026/09/15/inside-nvidias-cudnn-graph-api-fusion-autotuning-and-plan-reuse-with-cudnn-frontend"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmu37dt9m0b2urosatw9cn0g8","datePublished":"2026-09-15T21:37:11.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmu37dt9m0b2urosatw9cn0g8"}}],"aggregationSource":"MarkTechPost（RSS）","originalPublisher":{"name":"marktechpost.com","url":"https://www.marktechpost.com/2026/09/15/inside-nvidias-cudnn-graph-api-fusion-autotuning-and-plan-reuse-with-cudnn-frontend"},"geoDeepAnswer":null,"article":{"id":"cmu37dt9m0b2urosatw9cn0g8","slug":"cmu37dt9m0b2urosatw9cn0g8","url":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","title":"NVIDIA cuDNN Graph API 教程：融合、自动调优与 plan 复用","title_en":"","summary":"一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。","source":"MarkTechPost（RSS）","sourceUrl":"https://www.marktechpost.com/2026/09/15/inside-nvidias-cudnn-graph-api-fusion-autotuning-and-plan-reuse-with-cudnn-frontend","aiHotUrl":"https://aihot.news/items/cmu37dt9m0b2urosatw9cn0g8","publishedAt":"2026-09-15T21:37:11.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["In this tutorial：https://github.com/MARKTECHPOST-AI-MEDIA-INC/AI-Agents-Projects-Tutorials/blob/main/Deep%20Learning/cudnn_frontend_nvidia_tutorial_Marktechpost.ipynb , we work through the cuDNN Frontend ：https://github.com/NVIDIA/cudnn-frontend‘s graph API from below the framework: we describe a computation as a graph of operations, let cuDNN pick an engine to run it, and then take control of that choice ourselves. Every kernel we build here is expressed the same way: we declare tensors by their dimensions and strides, chain operations onto them, run the five-step build pipeline of validate, build operation graph, create execution plans, check support, and build plans, and then execute against a variant pack of pointers. We run it all on a single Colab GPU, checking each result against a PyTorch reference so we can see both that the fusion is correct and what it costs. The topics build on each other, moving from a single fused convolution to autotuning across engine configs, FP8-style epilogues, attention, plan serialization, dynamic shapes, and CUDA graph capture.","We start by installing nvidia-cudnn-frontend and solving the problem that trips up most first runs: making libcudnn.so visible to the frontend’s dynamic loader. We force PyTorch to load its bundled cuDNN first and then preload the shared objects explicitly, so the frontend’s own dlopen resolves against a library already resident in the process. We then report the compute capability, pick bfloat16 or float16 accordingly, create the cuDNN handle, and define the helpers for tensor description, graph building, workspace allocation, and event-based benchmarking that the rest of the notebook reuses.","We build our first graph, a convolution followed by a bias add and a ReLU, all fused into a single kernel. We keep every tensor in channels_last because that is what gives cuDNN the NHWC strides its tensor-core engines want, and we pin the output dimensions and strides explicitly so the result is written back in the same layout. We validate the output against torch.nn.functional.conv2d, then benchmark the fused graph against PyTorch running the convolution and activation as separate kernels.","We rebuild the same convolution but stop trusting the heuristic, asking for plans from heuristic modes A, B, and FALLBACK and compiling all of them with build_plan_policy.ALL. We then walk the plan list, build each config, allocate its specific workspace, and time it with execute_plan_at_index, printing throughput and workspace size for every candidate. The spread between the fastest and slowest engine is the point of the exercise, because it tells us how much we gain by shipping an autotuned index instead of accepting the default pick.","We move to a batched matmul and hang a full epilogue off it: an alpha scale supplied as a pass-by-value host scalar, a bias add, an activation, and an AMAX reduction over the result. The AMAX in the same kernel is the pattern that FP8 training relies on, since it collects the scale factor for the next quantization step without a second pass over the output. We compare against a PyTorch chain of baddbmm, activation, and amax, which makes clear that the speedup comes from eliminating epilogue memory traffic rather than from a faster GEMM.","We build a fused scaled dot-product attention graph with causal masking and check it against torch.nn.functional.scaled_dot_product_attention, guarding the whole section behind an SM80 check because the fused kernels need Ampere or newer. We write the causal argument with fallbacks, since the frontend has moved from use_causal_mask toward diagonal_alignment and bound arguments across its 1.x releases. We then serialize a built matmul graph to bytes, reload it into a fresh graph object, and execute it via integer UIDs, which lets us skip the compilation cost entirely at process startup.","We finish with two production concerns. First, we share a kernel cache across four graphs that differ only in batch size and time each build, so we can see later shapes reuse an already compiled kernel instead of paying the JIT cost again. Then we capture the convolution plan inside a CUDA graph, setting the cuDNN handle’s stream to the capture stream. So the work lands in the graph, and we measure how much per-iteration launch overhead the replay removes.","In conclusion, what we built here was small in code but broad in scope: a convolution, a matmul, and an attention kernel, each expressed as a graph rather than a library call. Working at that level changed what we could decide. We chose which operations collapsed into a single kernel, so the bias adds, activations, and AMAX reductions we folded into the epilogues never wrote an intermediate to memory. We chose the engine ourselves instead of accepting a heuristic, and timing every candidate config told us what that choice was worth. We also chose when to pay for compilation, pushing it out of the hot path with serialized plans, a kernel cache shared across shapes, and CUDA graph capture. The checks against PyTorch mattered as much as the timings, since the places where we merely matched it were usually places where PyTorch was already calling cuDNN underneath. That marked out where this API earns its keep: fusions with no framework-level equivalent, shapes hot enough to justify autotuning, and small kernels where startup and launch costs dominate.","Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58","Sana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.","Practitioner-first AI/ML news and analysis, read by 1M+ developers and researchers every month."],"articleImages":[{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/09/blog123-15-100x70.png","alt":"Meta Introduces ZGateway","afterParagraph":9,"url":"/media/articles/cmu37dt9m0b2urosatw9cn0g8/e9c71a1301d7b300.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/09/blog123-11-100x70.png","alt":"Sakana AI Researchers Introduce PC-ALM","afterParagraph":9,"url":"/media/articles/cmu37dt9m0b2urosatw9cn0g8/1f587fb7ee533826.png"}],"mediaStatus":"ok","articleBodyZh":["在本教程中：https://github.com/MARKTECHPOST-AI-MEDIA-INC/AI-Agents-Projects-Tutorials/blob/main/Deep%20Learning/cudnn_frontend_nvidia_tutorial_Marktechpost.ipynb，我们使用 cuDNN Frontend：https://github.com/NVIDIA/cudnn-frontend 的图 API，从框架下面进行操作：我们将计算描述为操作图，让 cuDNN 选择一个引擎来运行，然后自己控制这个选择。我们这里构建的每个内核都是以相同的方式表示的：我们按照张量的维度和步幅声明张量，将操作链式添加到它们上，运行五步构建流程：验证、构建操作图、创建执行计划、检查支持以及构建计划，然后针对指针变体包执行。我们在单个 Colab GPU 上运行所有操作，将每个结果与 PyTorch 的参考结果进行对比，以便同时看到融合是否正确以及其成本。各个主题相互构建，从单个融合卷积开始，到跨引擎配置的自动调优、FP8 风格的后处理、注意力机制、计划序列化、动态形状以及 CUDA 图捕获。","我们首先安装 nvidia-cudnn-frontend 并解决大多数首次运行时遇到的问题：使 libcudnn.so 对 frontend 的动态加载器可见。我们强制 PyTorch 先加载其捆绑的 cuDNN，然后显式预加载共享对象，以便 frontend 自身的 dlopen 可以针对已驻留在进程中的库解析。然后我们报告计算能力，据此选择 bfloat16 或 float16，创建 cuDNN 句柄，并定义张量描述、图构建、工作区分配以及基于事件的基准测试的辅助函数，这些函数将在整个笔记本中重复使用。","我们构建了第一个图，一个卷积后跟偏置加法和 ReLU，全部融合成一个内核。我们保持每个张量为 channels_last，因为这能为 cuDNN 提供其张量核心引擎所需的 NHWC 步幅，并且我们显式固定输出的维度和步幅，以便结果以相同布局写回。我们将输出与 torch.nn.functional.conv2d 进行验证，然后将融合图与 PyTorch 在单独的内核中运行卷积和激活操作的性能进行基准测试。","我们重新构建相同的卷积，但停止信任启发式方法，而是从启发式模式 A、B 和 FALLBACK 请求计划，并使用 build_plan_policy.ALL 编译它们。然后我们遍历计划列表，构建每个配置，分配其特定的工作空间，并使用 execute_plan_at_index 计时，打印每个候选的吞吐量和工作空间大小。最快和最慢引擎之间的差距就是这一实验的重点，因为它告诉我们通过发布自动调优索引相比接受默认选项能获得多少收益。","我们转向批量矩阵乘法，并在其上挂载完整的尾部操作：alpha 缩放（作为按值传递的主机标量提供）、偏置加、激活函数以及结果的 AMAX 归约。在同一个内核中进行 AMAX 是 FP8 训练依赖的模式，因为它在不需要对输出进行第二次遍历的情况下收集下一步量化的缩放因子。我们将其与 PyTorch 的 baddbmm、激活函数和 amax 链进行比较，这清楚地表明加速来自于消除了尾部操作的内存传输，而不是更快的 GEMM。","我们构建了带因果掩码的融合缩放点积注意力图，并将其与 torch.nn.functional.scaled_dot_product_attention 进行检查，整个部分的操作都受到 SM80 检查的保护，因为融合内核需要使用 Ampere 或更新版本。我们使用回退方式处理因果参数，因为前端已经从 use_causal_mask 转向 diagonal_alignment，并在 1.x 版本中绑定参数。然后我们将构建好的矩阵乘法图序列化为字节，重新加载到一个新的图对象中，并通过整数 UID 执行，这使我们可以在进程启动时完全跳过编译成本。","我们以两个生产相关问题结束。首先，我们在四个仅批量大小不同的图之间共享内核缓存，并计时每次构建，这样我们可以看到后续形状重用已经编译好的内核，而不必再次支付 JIT 成本。然后我们在 CUDA 图中捕获卷积计划，将 cuDNN 句柄的流设置为捕获流。这样工作就落在图中，我们可以测量重放消除后的每次迭代启动开销。","总之，我们在这里构建的内容在代码上很小，但范围很广：一个卷积、一个矩阵乘法和一个注意力内核，每个都表示为图而不是库调用。在这个层次上工作改变了我们能够做出的决策。我们选择了哪些操作合并到单个内核中，因此我们折叠到尾部的偏置加法、激活函数和 AMAX 归约从未将中间结果写入内存。我们自己选择了引擎，而不是接受启发式，每次对候选配置进行计时都告诉我们这个选择的价值。我们还选择何时支付编译成本，通过序列化计划、跨形状共享的内核缓存以及 CUDA 图捕获将其推到热路径之外。与 PyTorch 的对比检查和计时同样重要，因为我们仅刚好匹配 PyTorch 的地方通常是 PyTorch 已经在底层调用 cuDNN 的地方。这标志着这个 API 的价值所在：没有框架层等价物的融合、足够热门以值得自动调优的形状，以及启动和启动成本占主导的小内核。","需要与我们合作以推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？联系地址：https://forms.gle/wbash1wF6efRj8G58","Sana Hassan，是 Marktechpost 的咨询实习生，同时是 IIT 马德拉斯的双学位学生，热衷于应用技术和人工智能解决现实世界的挑战。由于对解决实际问题的浓厚兴趣，他为 AI 与现实解决方案的结合带来了新的视角。","以实践者为先的 AI/ML 新闻与分析，每月有超过 100 万开发者和研究人员阅读。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：公司与行业类动态需要放在竞争格局、商业化路径、资本信号和监管环境中观察，单条公告不能代表最终结果。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文件、合作落地、收入或用户信号、竞品动作和监管后续。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-09-18T17:13:51.592Z","sourceHash":"0758c71e243d3190","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["行业动态","MarkTechPost（RSS）"],"translations":{"zh-CN":{"title":"NVIDIA cuDNN Graph API 教程：融合、自动调优与 plan 复用","summary":"一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。","category":"行业动态","source":"marktechpost.com","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API 教程：融合、自动调优与 plan 复用 - Aioga AI资讯","description":"一篇教程详解 NVIDIA cuDNN Frontend 的 graph API：把计算声明为算子图，由 cuDNN 选择执行引擎，再手动接管该选择。","url":"https://www.aioga.com/news/cmu37dt9m0b2urosatw9cn0g8/","articleBody":["在本教程中：https://github.com/MARKTECHPOST-AI-MEDIA-INC/AI-Agents-Projects-Tutorials/blob/main/Deep%20Learning/cudnn_frontend_nvidia_tutorial_Marktechpost.ipynb，我们使用 cuDNN Frontend：https://github.com/NVIDIA/cudnn-frontend 的图 API，从框架下面进行操作：我们将计算描述为操作图，让 cuDNN 选择一个引擎来运行，然后自己控制这个选择。我们这里构建的每个内核都是以相同的方式表示的：我们按照张量的维度和步幅声明张量，将操作链式添加到它们上，运行五步构建流程：验证、构建操作图、创建执行计划、检查支持以及构建计划，然后针对指针变体包执行。我们在单个 Colab GPU 上运行所有操作，将每个结果与 PyTorch 的参考结果进行对比，以便同时看到融合是否正确以及其成本。各个主题相互构建，从单个融合卷积开始，到跨引擎配置的自动调优、FP8 风格的后处理、注意力机制、计划序列化、动态形状以及 CUDA 图捕获。","我们首先安装 nvidia-cudnn-frontend 并解决大多数首次运行时遇到的问题：使 libcudnn.so 对 frontend 的动态加载器可见。我们强制 PyTorch 先加载其捆绑的 cuDNN，然后显式预加载共享对象，以便 frontend 自身的 dlopen 可以针对已驻留在进程中的库解析。然后我们报告计算能力，据此选择 bfloat16 或 float16，创建 cuDNN 句柄，并定义张量描述、图构建、工作区分配以及基于事件的基准测试的辅助函数，这些函数将在整个笔记本中重复使用。","我们构建了第一个图，一个卷积后跟偏置加法和 ReLU，全部融合成一个内核。我们保持每个张量为 channels_last，因为这能为 cuDNN 提供其张量核心引擎所需的 NHWC 步幅，并且我们显式固定输出的维度和步幅，以便结果以相同布局写回。我们将输出与 torch.nn.functional.conv2d 进行验证，然后将融合图与 PyTorch 在单独的内核中运行卷积和激活操作的性能进行基准测试。","我们重新构建相同的卷积，但停止信任启发式方法，而是从启发式模式 A、B 和 FALLBACK 请求计划，并使用 build_plan_policy.ALL 编译它们。然后我们遍历计划列表，构建每个配置，分配其特定的工作空间，并使用 execute_plan_at_index 计时，打印每个候选的吞吐量和工作空间大小。最快和最慢引擎之间的差距就是这一实验的重点，因为它告诉我们通过发布自动调优索引相比接受默认选项能获得多少收益。","我们转向批量矩阵乘法，并在其上挂载完整的尾部操作：alpha 缩放（作为按值传递的主机标量提供）、偏置加、激活函数以及结果的 AMAX 归约。在同一个内核中进行 AMAX 是 FP8 训练依赖的模式，因为它在不需要对输出进行第二次遍历的情况下收集下一步量化的缩放因子。我们将其与 PyTorch 的 baddbmm、激活函数和 amax 链进行比较，这清楚地表明加速来自于消除了尾部操作的内存传输，而不是更快的 GEMM。","我们构建了带因果掩码的融合缩放点积注意力图，并将其与 torch.nn.functional.scaled_dot_product_attention 进行检查，整个部分的操作都受到 SM80 检查的保护，因为融合内核需要使用 Ampere 或更新版本。我们使用回退方式处理因果参数，因为前端已经从 use_causal_mask 转向 diagonal_alignment，并在 1.x 版本中绑定参数。然后我们将构建好的矩阵乘法图序列化为字节，重新加载到一个新的图对象中，并通过整数 UID 执行，这使我们可以在进程启动时完全跳过编译成本。","我们以两个生产相关问题结束。首先，我们在四个仅批量大小不同的图之间共享内核缓存，并计时每次构建，这样我们可以看到后续形状重用已经编译好的内核，而不必再次支付 JIT 成本。然后我们在 CUDA 图中捕获卷积计划，将 cuDNN 句柄的流设置为捕获流。这样工作就落在图中，我们可以测量重放消除后的每次迭代启动开销。","总之，我们在这里构建的内容在代码上很小，但范围很广：一个卷积、一个矩阵乘法和一个注意力内核，每个都表示为图而不是库调用。在这个层次上工作改变了我们能够做出的决策。我们选择了哪些操作合并到单个内核中，因此我们折叠到尾部的偏置加法、激活函数和 AMAX 归约从未将中间结果写入内存。我们自己选择了引擎，而不是接受启发式，每次对候选配置进行计时都告诉我们这个选择的价值。我们还选择何时支付编译成本，通过序列化计划、跨形状共享的内核缓存以及 CUDA 图捕获将其推到热路径之外。与 PyTorch 的对比检查和计时同样重要，因为我们仅刚好匹配 PyTorch 的地方通常是 PyTorch 已经在底层调用 cuDNN 的地方。这标志着这个 API 的价值所在：没有框架层等价物的融合、足够热门以值得自动调优的形状，以及启动和启动成本占主导的小内核。","需要与我们合作以推广您的 GitHub 仓库或 Hugging Face 页面或产品发布或网络研讨会等吗？联系地址：https://forms.gle/wbash1wF6efRj8G58","Sana Hassan，是 Marktechpost 的咨询实习生，同时是 IIT 马德拉斯的双学位学生，热衷于应用技术和人工智能解决现实世界的挑战。由于对解决实际问题的浓厚兴趣，他为 AI 与现实解决方案的结合带来了新的视角。","以实践者为先的 AI/ML 新闻与分析，每月有超过 100 万开发者和研究人员阅读。"]},"en":{"title":"NVIDIA cuDNN Graph API Tutorial: Fusion, Auto-Tuning, and Plan Reuse","summary":"A tutorial detailing NVIDIA cuDNN Frontend's graph API: declare computations as an operator graph, let cuDNN choose the execution engine, and then manually take over that choice.","category":"Industry","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API Tutorial: Fusion, Auto-Tuning, and Plan Reuse - Aioga AI News","description":"A tutorial detailing NVIDIA cuDNN Frontend's graph API: declare computations as an operator graph, let cuDNN choose the execution engine, and then manually take over that choice.","url":"https://www.aioga.com/en/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:06:12.142Z"},"ja":{"title":"NVIDIA cuDNN Graph API チュートリアル：融合、オートチューニングとプラン再利用","summary":"NVIDIA cuDNN Frontend の graph API を詳しく解説するチュートリアル：計算をオペレーターグラフとして宣言し、cuDNN に実行エンジンを選ばせた後、その選択を手動で引き継ぐ。","category":"業界動向","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API チュートリアル：融合、オートチューニングとプラン再利用 - Aioga AIニュース","description":"NVIDIA cuDNN Frontend の graph API を詳しく解説するチュートリアル：計算をオペレーターグラフとして宣言し、cuDNN に実行エンジンを選ばせた後、その選択を手動で引き継ぐ。","url":"https://www.aioga.com/ja/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:06:14.649Z"},"ko":{"title":"NVIDIA cuDNN Graph API 튜토리얼: 융합, 자동 최적화 및 계획 재사용","summary":"한 튜토리얼에서 NVIDIA cuDNN Frontend의 graph API를 상세히 설명합니다: 계산을 연산자 그래프로 선언하고, cuDNN이 실행 엔진을 선택하게 한 다음, 그 선택을 수동으로 인계받습니다.","category":"업계 동향","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API 튜토리얼: 융합, 자동 최적화 및 계획 재사용 - Aioga AI 뉴스","description":"한 튜토리얼에서 NVIDIA cuDNN Frontend의 graph API를 상세히 설명합니다: 계산을 연산자 그래프로 선언하고, cuDNN이 실행 엔진을 선택하게 한 다음, 그 선택을 수동으로 인계받습니다.","url":"https://www.aioga.com/ko/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:07:07.011Z"},"es":{"title":"Tutorial de NVIDIA cuDNN Graph API: fusión, ajuste automático y reutilización de planes","summary":"Un tutorial detalla la API de gráficos de NVIDIA cuDNN Frontend: declarar los cálculos como un gráfico de operadores, dejar que cuDNN elija el motor de ejecución, y luego tomar el control manualmente de esa elección.","category":"Industria","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Tutorial de NVIDIA cuDNN Graph API: fusión, ajuste automático y reutilización de planes - Aioga Noticias de IA","description":"Un tutorial detalla la API de gráficos de NVIDIA cuDNN Frontend: declarar los cálculos como un gráfico de operadores, dejar que cuDNN elija el motor de ejecución, y luego tomar el...","url":"https://www.aioga.com/es/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:06:57.161Z"},"fr":{"title":"Tutoriel sur l'API Graph de NVIDIA cuDNN : fusion, optimisation automatique et réutilisation de plan","summary":"Un tutoriel détaillé expliquant l'API graph de NVIDIA cuDNN Frontend : déclarer les calculs sous forme de graphe d'opérateurs, laisser cuDNN choisir le moteur d'exécution, puis reprendre manuellement ce choix.","category":"Industrie","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Tutoriel sur l'API Graph de NVIDIA cuDNN : fusion, optimisation automatique et réutilisation de plan - Aioga Actualités IA","description":"Un tutoriel détaillé expliquant l'API graph de NVIDIA cuDNN Frontend : déclarer les calculs sous forme de graphe d'opérateurs, laisser cuDNN choisir le moteur d'exécution, puis rep...","url":"https://www.aioga.com/fr/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:07:53.461Z"},"de":{"title":"NVIDIA cuDNN Graph API Tutorial: Fusion, automatische Optimierung und Plan-Wiederverwendung","summary":"Ein Tutorial erklärt detailliert die Graph-API von NVIDIA cuDNN Frontend: Deklariert Berechnungen als Operator-Graph, lässt cuDNN die Ausführungsengine auswählen und übernimmt dann diese Auswahl manuell.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API Tutorial: Fusion, automatische Optimierung und Plan-Wiederverwendung - Aioga KI-News","description":"Ein Tutorial erklärt detailliert die Graph-API von NVIDIA cuDNN Frontend: Deklariert Berechnungen als Operator-Graph, lässt cuDNN die Ausführungsengine auswählen und übernimmt dann...","url":"https://www.aioga.com/de/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:07:54.378Z"},"pt-BR":{"title":"Tutorial da API NVIDIA cuDNN Graph: fusão, ajuste automático e reutilização de plano","summary":"Um tutorial detalhado sobre a API de gráfico do NVIDIA cuDNN Frontend: declarar os cálculos como um gráfico de operadores, deixar o cuDNN escolher o mecanismo de execução e, em seguida, assumir manualmente essa escolha.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Tutorial da API NVIDIA cuDNN Graph: fusão, ajuste automático e reutilização de plano - Aioga Notícias de IA","description":"Um tutorial detalhado sobre a API de gráfico do NVIDIA cuDNN Frontend: declarar os cálculos como um gráfico de operadores, deixar o cuDNN escolher o mecanismo de execução e, em seg...","url":"https://www.aioga.com/pt-BR/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:08:35.317Z"},"ru":{"title":"Учебник по API графов NVIDIA cuDNN: слияние, автоматическая настройка и повторное использование планов","summary":"Подробное руководство по графическому API NVIDIA cuDNN Frontend: объявление вычислений в виде графа операторов, выбор исполнительного движка cuDNN, а затем ручное управление этим выбором.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Учебник по API графов NVIDIA cuDNN: слияние, автоматическая настройка и повторное использование планов - Aioga Новости ИИ","description":"Подробное руководство по графическому API NVIDIA cuDNN Frontend: объявление вычислений в виде графа операторов, выбор исполнительного движка cuDNN, а затем ручное управление этим в...","url":"https://www.aioga.com/ru/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:08:38.291Z"},"ar":{"title":"دليل واجهة برمجة تطبيقات NVIDIA cuDNN Graph: الدمج، التحسين التلقائي وإعادة استخدام الخطة","summary":"دليل تفصيلي يشرح واجهة برمجة تطبيقات الرسوميات (graph API) لـ NVIDIA cuDNN Frontend: تحويل الحساب إلى رسم بياني للعمليات، يختار cuDNN محرك التنفيذ، ثم تتولى التحكم في هذا الاختيار يدويًا.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"دليل واجهة برمجة تطبيقات NVIDIA cuDNN Graph: الدمج، التحسين التلقائي وإعادة استخدام الخطة - Aioga أخبار الذكاء الاصطناعي","description":"دليل تفصيلي يشرح واجهة برمجة تطبيقات الرسوميات (graph API) لـ NVIDIA cuDNN Frontend: تحويل الحساب إلى رسم بياني للعمليات، يختار cuDNN محرك التنفيذ، ثم تتولى التحكم في هذا الاختيار...","url":"https://www.aioga.com/ar/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:09:25.339Z"},"hi":{"title":"NVIDIA cuDNN ग्राफ़ API ट्यूटोरियल: फ्यूज़न, स्वचालित ट्यूनिंग और प्लान पुन: उपयोग","summary":"एक ट्यूटोरियल NVIDIA cuDNN Frontend के ग्राफ़ API की विस्तृत व्याख्या करता है: गणना को ऑपरेटर ग्राफ़ के रूप में घोषित करना, cuDNN द्वारा निष्पादन इंजन का चयन करना, और फिर उस चयन को मैन्युअल रूप से संभालना।","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN ग्राफ़ API ट्यूटोरियल: फ्यूज़न, स्वचालित ट्यूनिंग और प्लान पुन: उपयोग - Aioga AI समाचार","description":"एक ट्यूटोरियल NVIDIA cuDNN Frontend के ग्राफ़ API की विस्तृत व्याख्या करता है: गणना को ऑपरेटर ग्राफ़ के रूप में घोषित करना, cuDNN द्वारा निष्पादन इंजन का चयन करना, और फिर उस चयन को...","url":"https://www.aioga.com/hi/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:09:20.926Z"},"it":{"title":"Tutorial NVIDIA cuDNN Graph API: fusione, ottimizzazione automatica e riutilizzo del piano","summary":"Un tutorial che spiega in dettaglio l'API graph di NVIDIA cuDNN Frontend: dichiarare il calcolo come un grafo di operatori, lasciare che cuDNN scelga il motore di esecuzione, e poi prendere manualmente il controllo di tale scelta.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Tutorial NVIDIA cuDNN Graph API: fusione, ottimizzazione automatica e riutilizzo del piano - Aioga Notizie IA","description":"Un tutorial che spiega in dettaglio l'API graph di NVIDIA cuDNN Frontend: dichiarare il calcolo come un grafo di operatori, lasciare che cuDNN scelga il motore di esecuzione, e poi...","url":"https://www.aioga.com/it/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:10:08.553Z"},"nl":{"title":"NVIDIA cuDNN Graph API Tutorial: Fusie, automatische afstelling en planhergebruik","summary":"Een tutorial die de NVIDIA cuDNN Frontend graph API in detail uitlegt: verklaar berekeningen als een operatorgrafiek, laat cuDNN de uitvoeringsengine kiezen, en neem daarna handmatig die keuze over.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API Tutorial: Fusie, automatische afstelling en planhergebruik - Aioga AI-nieuws","description":"Een tutorial die de NVIDIA cuDNN Frontend graph API in detail uitlegt: verklaar berekeningen als een operatorgrafiek, laat cuDNN de uitvoeringsengine kiezen, en neem daarna handmat...","url":"https://www.aioga.com/nl/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:10:06.149Z"},"tr":{"title":"NVIDIA cuDNN Graph API Eğitimi: Birleştirme, Otomatik Optimizasyon ve Plan Tekrar Kullanımı","summary":"Bir kılavuz, NVIDIA cuDNN Frontend'in graph API'sini ayrıntılı olarak açıklıyor: Hesaplamayı bir operatör grafiği olarak bildirin, cuDNN yürütme motorunu seçsin, ardından bu seçimi manuel olarak devralın.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"NVIDIA cuDNN Graph API Eğitimi: Birleştirme, Otomatik Optimizasyon ve Plan Tekrar Kullanımı - Aioga AI Haberleri","description":"Bir kılavuz, NVIDIA cuDNN Frontend'in graph API'sini ayrıntılı olarak açıklıyor: Hesaplamayı bir operatör grafiği olarak bildirin, cuDNN yürütme motorunu seçsin, ardından bu seçimi...","url":"https://www.aioga.com/tr/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:10:50.914Z"},"vi":{"title":"Hướng dẫn API Đồ thị cuDNN của NVIDIA: Hợp nhất, tối ưu tự động và tái sử dụng plan","summary":"Một bài hướng dẫn chi tiết về API đồ thị của NVIDIA cuDNN Frontend: khai báo phép tính dưới dạng đồ thị toán tử, để cuDNN chọn engine thực thi, sau đó tự tay tiếp quản lựa chọn đó.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Hướng dẫn API Đồ thị cuDNN của NVIDIA: Hợp nhất, tối ưu tự động và tái sử dụng plan - Tin tức AI Aioga","description":"Một bài hướng dẫn chi tiết về API đồ thị của NVIDIA cuDNN Frontend: khai báo phép tính dưới dạng đồ thị toán tử, để cuDNN chọn engine thực thi, sau đó tự tay tiếp quản lựa chọn đó.","url":"https://www.aioga.com/vi/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:10:52.816Z"},"id":{"title":"Tutorial NVIDIA cuDNN Graph API: Fusi, Penyetelan Otomatis, dan Penggunaan Ulang Plan","summary":"Sebuah tutorial yang menjelaskan secara rinci API graph NVIDIA cuDNN Frontend: menyatakan perhitungan sebagai grafik operator, membiarkan cuDNN memilih mesin eksekusi, lalu mengambil alih pilihan tersebut secara manual.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Tutorial NVIDIA cuDNN Graph API: Fusi, Penyetelan Otomatis, dan Penggunaan Ulang Plan - Berita AI Aioga","description":"Sebuah tutorial yang menjelaskan secara rinci API graph NVIDIA cuDNN Frontend: menyatakan perhitungan sebagai grafik operator, membiarkan cuDNN memilih mesin eksekusi, lalu mengamb...","url":"https://www.aioga.com/id/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:11:35.720Z"},"th":{"title":"บทเรียน NVIDIA cuDNN Graph API: การรวม การปรับแต่งอัตโนมัติ และการนำแผนกลับมาใช้ใหม่","summary":"บทเรียนนี้อธิบายอย่างละเอียดเกี่ยวกับ NVIDIA cuDNN Frontend graph API: ประกาศการคำนวณเป็นกราฟของโอเปอเรเตอร์ ให้ cuDNN เลือกเอนจินสำหรับการประมวลผล จากนั้นจึงเข้าไปควบคุมการเลือกนั้นด้วยตนเอง","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"บทเรียน NVIDIA cuDNN Graph API: การรวม การปรับแต่งอัตโนมัติ และการนำแผนกลับมาใช้ใหม่ - ข่าว AI Aioga","description":"บทเรียนนี้อธิบายอย่างละเอียดเกี่ยวกับ NVIDIA cuDNN Frontend graph API: ประกาศการคำนวณเป็นกราฟของโอเปอเรเตอร์ ให้ cuDNN เลือกเอนจินสำหรับการประมวลผล จากนั้นจึงเข้าไปควบคุมการเลือกนั...","url":"https://www.aioga.com/th/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:11:41.449Z"},"pl":{"title":"Samouczek NVIDIA cuDNN Graph API: fuzja, automatyczna optymalizacja i ponowne użycie planu","summary":"Samouczek szczegółowo wyjaśniający API grafu frontendu NVIDIA cuDNN: deklarowanie obliczeń jako grafu operatorów, pozwalanie cuDNN na wybór silnika wykonawczego, a następnie ręczne przejęcie tego wyboru.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Samouczek NVIDIA cuDNN Graph API: fuzja, automatyczna optymalizacja i ponowne użycie planu - Aioga Wiadomości AI","description":"Samouczek szczegółowo wyjaśniający API grafu frontendu NVIDIA cuDNN: deklarowanie obliczeń jako grafu operatorów, pozwalanie cuDNN na wybór silnika wykonawczego, a następnie ręczne...","url":"https://www.aioga.com/pl/news/cmu37dt9m0b2urosatw9cn0g8/","contentTranslated":true,"sourceHash":"ae852ff939990957","translatedAt":"2026-09-15T23:12:25.684Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":""}}