[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-blackwell-mlperf-training-6-0-sweep-zh":3,"article-related-blackwell-mlperf-training-6-0-sweep-zh":35,"series-industry-2e953e03-6c74-468e-ab2f-9b698a3b2d39":83},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":27,"views":31,"created_at":32,"published_at":33,"topic_cluster_id":34},"2e953e03-6c74-468e-ab2f-9b698a3b2d39","blackwell-mlperf-training-6-0-sweep-zh","Blackwell 6.0 讓訓練速度、規模、穩定性一起升級","\u003Cp data-speakable=\"summary\">Blackwell 在 MLPerf Training 6.0 拿下七項最快成績，適合想同時提升訓練速度、規模與可靠性的 AI 團隊。\u003C\u002Fp>\u003Cp>看完這 5 項，你可以直接判斷 Blackwell 值不值得拿來做前沿模型訓練，尤其是當你在意的不只是單次跑分，而是能不能把更大的模型更快、也更穩地訓完。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>規模\u003C\u002Fth>\u003Cth>報告結果\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">NVIDIA Blackwell\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>MLPerf Training 6.0\u003C\u002Ftd>\u003Ctd>七項 benchmark 都是最快 time to train\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">GB300 NVL72\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>Rack-scale\u003C\u002Ftd>\u003Ctd>比 GB200 NVL72 最多快 1.6x\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">DeepSeek-V3 671B\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>8,192 GPUs\u003C\u002Ftd>\u003Ctd>最大規模下仍是最快訓練\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.microsoft.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Azure 上的 Llama 3.1 405B\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>8,192 GPUs\u003C\u002Ftd>\u003Ctd>7.07 分鐘達到 reference quality\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>\u003Ca href=\"https:\u002F\u002Fwww.coreweave.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">CoreWeave 上的 DeepSeek-V3 671B\u003C\u002Fa>\u003C\u002Ftd>\u003Ctd>8,192 GPUs\u003C\u002Ftd>\u003Ctd>2.02 分鐘達到 reference quality\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. 七項 benchmark 都是最快\u003C\u002Fh2>\u003Cp>這次最關鍵的\u003Ca href=\"\u002Fnews\u002Fai-market-map-list-better-signal-than-newsletters-zh\">訊號\u003C\u002Fa>，不是單一模型成績，而是 \u003Ca href=\"https:\u002F\u002Fmlcommons.org\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">MLPerf Training\u003C\u002Fa> 6.0 的整體表現。\u003Ca href=\"\u002Ftag\u002Fnvidia\">NVIDIA\u003C\u002Fa> 是唯一跨七個 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 都有提交的陣營，而且每一項都拿到最快 time to train。對訓練基礎設施來說，這代表它不是只對某一類模型特別有利，而是把「更快完成訓練」當成通用目標。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782181969967-chye.png\" alt=\"Blackwell 6.0 讓訓練速度、規模、穩定性一起升級\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種結果對採購和架構選型很直接。你如果在看 dense \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa>、MoE，或是 fine-tuning 工作流，這份成績單給的是同一個答案：Blackwell 不只是在某個 demo 上快，而是能在多種訓練型態裡維持領先。\u003C\u002Fp>\u003Cul>\u003Cli>7\u002F7 benchmark 都是最快\u003C\u002Fli>\u003Cli>同時涵蓋 GB200 NVL72 與 GB300 NVL72\u003C\u002Fli>\u003Cli>包含新的 MoE 工作負載，如 DeepSeek-V3 671B、GPT-OSS-20B\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. GB300 NVL72 把單機櫃速度再往上推\u003C\u002Fh2>\u003Cp>如果你關心的是同樣 rack-scale 架構下能再省多少時間，GB300 NVL72 是最值得看的結果。NVIDIA 表示，它在相同規模下比 GB200 NVL72 最多快 1.6x，原因包括更高的 compute density、更多記憶體，以及更高的功耗上限，讓效能可以更長時間維持在高位。\u003C\u002Fp>\u003Cp>這種提升對長時間 pretraining 或反覆 fine-tune 特別有感。當模型本來就很大時，1.6x 不只是分數好看，而是直接改變排程，原本要跨好幾天的工作，可能就能少掉一大段等待時間。\u003C\u002Fp>\u003Ccode>GB300 NVL72 的主要增益來源：\n- 更高 compute density，搭配 NVFP4\n- 更大的記憶體容量\n- 更高功耗上限，支撐持續輸出\u003C\u002Fcode>\u003Ch2>3. 8,192 GPU 規模證明它不是只會跑小實驗\u003C\u002Fh2>\u003Cp>Blackwell 的另一個重點是能不能把規模做大，而且還維持效率。NVIDIA 把 DeepSeek-V3 671B 擴到 8,192 GPUs，這是目前 MLPerf Training 裡最大的 Blackwell 相關提交。它也把 Llama 3.1 405B 跑到 5,120 GPUs，顯示平台不只是拚單點速度，也是真的能把叢集往外擴。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782181966040-2v3f.png\" alt=\"Blackwell 6.0 讓訓練速度、規模、穩定性一起升級\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>能撐住這種規模，靠的是互連設計。每個 rack 內，\u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">第五代 NVLink Switch\u003C\u002Fa> 把 72 顆 GPU 連成共享的運算與記憶體池；跨機櫃則依\u003Ca href=\"\u002Ftag\u002F資料中心\">資料中心\u003C\u002Fa>設計，搭配 \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Quantum InfiniBand\u003C\u002Fa> 或 \u003Ca href=\"https:\u002F\u002Fwww.nvidia.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Spectrum-X Ethernet\u003C\u002Fa>。\u003C\u002Fp>\u003Cul>\u003Cli>DeepSeek-V3 671B：8,192 GPUs\u003C\u002Fli>\u003Cli>Llama 3.1 405B：5,120 GPUs\u003C\u002Fli>\u003Cli>單個 rack 內 72 顆 GPU 共享互連\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>4. 合作案例顯示它已經進到真實產線\u003C\u002Fh2>\u003Cp>只看 NVIDIA 自家成績還不夠，合作夥伴案例更能看出平台是不是已經進到 production。\u003Ca href=\"https:\u002F\u002Fcohere.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Cohere\u003C\u002Fa> 表示，它在 GB200 NVL72 上讓 North \u003Ca href=\"\u002Ftag\u002Fagentic-ai\">agentic AI\u003C\u002Fa> platform 的訓練速度快了 3 倍；\u003Ca href=\"https:\u002F\u002Fwww.midjourney.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Midjourney\u003C\u002Fa> 則在 Blackwell 叢集上訓練 v8，並在 \u003Ca href=\"https:\u002F\u002Fwww.coreweave.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">CoreWeave\u003C\u002Fa> 上擴大 Blackwell Ultra GPU fleet，準備下一代影像與影片模型。\u003C\u002Fp>\u003Cp>其他案例也很有說服力。\u003Ca href=\"https:\u002F\u002Fazure.microsoft.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Microsoft Azure\u003C\u002Fa> 在 Llama 3.1 405B 上 7.07 分鐘達到 reference quality；CoreWeave 在 DeepSeek-V3 671B 上用 GB300 NVL72 只花 2.02 分鐘；\u003Ca href=\"https:\u002F\u002Fwww.nebius.com\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Nebius\u003C\u002Fa> 則說 \u003Ca href=\"https:\u002F\u002Fwww.higgsfield.ai\u002F\" target=\"_blank\" rel=\"noopener noreferrer\">Higgsfield\u003C\u002Fa> 的訓練時間縮短了 30%。\u003C\u002Fp>\u003Cul>\u003Cli>Cohere：GB200 NVL72 上快 3 倍\u003C\u002Fli>\u003Cli>Midjourney：已在 Blackwell 上訓練並擴張\u003C\u002Fli>\u003Cli>Azure、CoreWeave、Nebius 都有可量化成果\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. 長訓練最怕中斷，Blackwell 把這點補上\u003C\u002Fh2>\u003Cp>訓練速度再快，如果跑到一半出錯，整體成本還是會爆。NVIDIA 把 Blackwell 的可靠性拆成幾層：出廠前要經過 30+ 道\u003Ca href=\"\u002Fnews\u002Fworldcoin-rally-credibility-test-not-breakout-zh\">測試\u003C\u002Fa>，運作時由 Reliability, Availability and Serviceability Engine 幾乎全芯片監控，遇到故障還能用 self-healing 邏輯繞開問題點。\u003C\u002Fp>\u003Cp>在叢集層，Spectrum-X Ethernet 可以在毫秒內改道失效連線；如果真的中斷，NVIDIA Resiliency Extension，簡稱 NVRx，可以從最近 checkpoint 接著跑，而不是整個重來。對動輒數週、數月、數十萬張 GPU 的工作來說，這些功能往往比單純的峰值效能更值錢。\u003C\u002Fp>\u003Ccode>可靠性堆疊：\n- 30+ 道出廠測試\n- RAS Engine 監控\n- self-healing fault routing\n- Spectrum-X 毫秒級改道\n- NVRx checkpoint 恢復\u003C\u002Fcode>\u003Ch2>怎麼挑\u003C\u002Fh2>\u003Cp>如果你最在意 benchmark 成績，先看七項全勝和 GB300 NVL72 的 1.6x 提升；如果你需要把模型做大，8,192 GPUs 的 DeepSeek-V3 671B 是最強證據；如果你想知道它是否已經進入\u003Ca href=\"\u002Fnews\u002F5-2026-new-crypto-picks-utility-zh\">真實\u003C\u002Fa>業務，Cohere、Midjourney、Azure、CoreWeave 和 Nebius 的案例更有參考價值。\u003C\u002Fp>\u003Cp>整體來看，Blackwell 比較像一個完整訓練平台，而不是單顆很快的 GPU。適合前沿模型團隊、長週期 pretraining 團隊，以及不能接受頻繁重跑的工作負載。\u003C\u002Fp>","5 項 Blackwell MLPerf 6.0 成績，幫你判斷它適不適合追求更快訓練、更大規模與更高可靠性的團隊。","blogs.nvidia.com","https:\u002F\u002Fblogs.nvidia.com\u002Fblog\u002Fblackwell-mlperf-training-6-0\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782181969967-chye.png","industry","zh","50270b29-033a-468f-a188-9163b77e0d0c",[17,18,19,20,21,22,23,24,25,26],"Blackwell","MLPerf Training 6.0","NVIDIA","GB300 NVL72","GB200 NVL72","DeepSeek-V3 671B","Llama 3.1 405B","可靠性","GPU 叢集","AI 訓練",[28,29,30],"Blackwell 在 MLPerf Training 6.0 七項 benchmark 都拿到最快 time to train。","GB300 NVL72 相比 GB200 NVL72 最多快 1.6x，適合長訓練工作。","8,192 GPU 規模與多家夥伴案例顯示，它已能支撐真實產線。",0,"2026-06-23T02:32:25.659395+00:00","2026-06-23T02:32:25.648+00:00","3ae0b1d4-3133-42fa-8d6d-af2d36c8b11f",{"tags":36,"relatedLang":42,"relatedPosts":46},[37,40],{"name":38,"slug":39},"Nvidia","nvidia",{"name":17,"slug":41},"blackwell",{"id":15,"slug":43,"title":44,"language":45},"blackwell-mlperf-training-6-0-sweep-en","Blackwell’s MLPerf sweep shows why training speeds up","en",[47,53,59,65,71,77],{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"2166c4ec-e128-4281-931e-30b896b6c0ca","baya-openchip-bet-ai-silicon-data-movement-zh","Baya 與 Openchip 押注 AI 晶片未來：關鍵不是算力，而是資料搬運","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782193668588-hvs5.png","2026-06-23T05:47:23.424804+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"031360c4-4d0a-46bf-bbd1-ec2db9392537","citigroup-sees-tokenized-assets-hitting-8-2t-zh","Citi估8.2兆美元資產將被代幣化","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782185577271-2wur.png","2026-06-23T03:32:33.388896+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"9bff47f9-faba-408a-b0f9-83d050ed3ab6","rwa-tokenization-turns-assets-into-on-chain-rails-zh","RWA 代幣化把資產變上鏈通道","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782184696459-t7qe.png","2026-06-23T03:17:50.080174+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"0bcd7d43-04e9-455e-b0c0-a9d9679a50a2","micron-anthropic-deal-earnings-signal-zh","Micron 合約把財報變訊號","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782183794288-6y4p.png","2026-06-23T03:02:49.715904+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"039f17cc-fbd9-4c4a-9427-952af4c962ec","ai-companies-should-stop-pretending-midterm-spending-is-neut-zh","AI 公司該停止把中期選舉支出說成中立","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782181064576-t6xw.png","2026-06-23T02:17:18.930817+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":13},"ecd1f521-2d1a-411a-a7c0-bcd68193a21b","ai-market-map-list-better-signal-than-newsletters-zh","這份 AI 市場地圖清單，比大多數 AI 電子報更有訊號","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782180167709-2byc.png","2026-06-23T02:02:21.657988+00:00",[84,89,94,99,104,109,114,119,124,129],{"id":85,"slug":86,"title":87,"created_at":88},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]