[{"data":1,"prerenderedAt":759},["ShallowReactive",2],{"content-/amd-mi355x-vs-nvidia-b200-cost-analysis":3,"all-pages-for-dir":737,"related-/amd-mi355x-vs-nvidia-b200-cost-analysis":738,"og-image-/amd-mi355x-vs-nvidia-b200-cost-analysis":758},{"id":4,"title":5,"body":6,"category":720,"concepts":720,"description":721,"extension":722,"meta":723,"navigation":363,"ogImage":720,"path":724,"project_name":720,"published":725,"publishedAt":726,"seo":727,"source":720,"stem":728,"tags":729,"todo":720,"unpublished":725,"updatedAt":720,"__hash__":736},"pages/2025-10/2025-10-17/amd-mi355x-vs-nvidia-b200-cost-analysis.md","AMD MI355X vs NVIDIA B200 - 1百万トークンあたりコストで見る真の勝者",{"type":7,"value":8,"toc":701},"minimark",[9,14,19,24,28,36,39,43,132,139,141,145,148,167,170,172,176,180,216,219,222,224,228,280,287,289,293,298,505,509,520,524,536,540,548,550,554,623,625,629,649,652,658,660,664,695,697],[10,11,13],"h1",{"id":12},"amd-mi355x-vs-nvidia-b200","AMD MI355X vs NVIDIA B200",[15,16,18],"h2",{"id":17},"_1百万トークンあたりコストで見る真の勝者はどちらか","―「1百万トークンあたりコスト」で見る真の勝者はどちらか",[20,21,23],"h3",{"id":22},"はじめに","🔍 はじめに",[25,26,27],"p",{},"2025年のAIインフラ競争は、性能よりもコスト効率（TCO / Cost per Million Tokens）が主戦場になりつつあります。\n最近のSemiAnalysisレポートでは、AMDのMI355XがNVIDIAのB200を一部条件下で上回るという主張がありましたが、これは本当に現実的な比較なのでしょうか。",[25,29,30,31,35],{},"ここでは、公開データ・実測・仮定モデルをもとに、両者の",[32,33,34],"strong",{},"実効コスト","を徹底比較します。",[37,38],"hr",{},[15,40,42],{"id":41},"_1-基本スペック比較","🧩 1. 基本スペック比較",[44,45,46,62],"table",{},[47,48,49],"thead",{},[50,51,52,56,59],"tr",{},[53,54,55],"th",{},"項目",[53,57,58],{},"AMD MI355X",[53,60,61],{},"NVIDIA B200",[63,64,65,77,88,99,110,121],"tbody",{},[50,66,67,71,74],{},[68,69,70],"td",{},"アーキテクチャ",[68,72,73],{},"CDNA4",[68,75,76],{},"Blackwell",[50,78,79,82,85],{},[68,80,81],{},"メモリ",[68,83,84],{},"HBM3E 288 GB / 8 TB/s",[68,86,87],{},"HBM3E 約192 GB / ~7.7 TB/s",[50,89,90,93,96],{},[68,91,92],{},"TDP",[68,94,95],{},"約1,400 W（液冷）",[68,97,98],{},"約1,787 W（DGX B200換算）",[50,100,101,104,107],{},[68,102,103],{},"FP8演算性能",[68,105,106],{},"約10 PFLOPS",[68,108,109],{},"約8 〜 10 PFLOPS（構成依存）",[50,111,112,115,118],{},[68,113,114],{},"FP4演算性能",[68,116,117],{},"約20 PFLOPS",[68,119,120],{},"約18 〜 20 PFLOPS",[50,122,123,126,129],{},[68,124,125],{},"システム構成",[68,127,128],{},"Infinity Fabric ×8 GPU",[68,130,131],{},"NVLink/NVSwitch ×72 GPU（NVL72）",[25,133,134,135,138],{},"💡",[32,136,137],{},"ポイント","：\nMI355Xはメモリ容量とFP4/FP6性能で優位。\n一方、B200はNVSwitchスケール性能・ソフトウェア最適化で優勢です。",[37,140],{},[15,142,144],{"id":143},"️-2-コスト構造の前提","⚙️ 2. コスト構造の前提",[25,146,147],{},"TCO（総所有コスト）は、以下の変数で大きく変わります。",[149,150,151,155,158,161,164],"ul",{},[152,153,154],"li",{},"GPU本体価格（AMDはNVIDIA比約30％安いと推定）",[152,156,157],{},"消費電力＋冷却インフラコスト（OPEX）",[152,159,160],{},"実効性能（スループット／トークン処理速度）",[152,162,163],{},"ソフト最適化・サポート費用",[152,165,166],{},"利用率・契約形態（オンデマンド／リザーブド等）",[25,168,169],{},"こうした複合要素のため、TCOは見かけ上AMDが安く見える構造になりがちです。",[37,171],{},[15,173,175],{"id":174},"_3-1百万トークンあたりコスト比較","💰 3. 「1百万トークンあたりコスト」比較",[20,177,179],{"id":178},"_31-tcoモデル上の比較amd公称","3.1 TCOモデル上の比較（AMD公称）",[44,181,182,198],{},[47,183,184],{},[50,185,186,189,192,195],{},[53,187,188],{},"指標",[53,190,191],{},"MI355X",[53,193,194],{},"B200",[53,196,197],{},"差",[63,199,200],{},[50,201,202,205,208,211],{},[68,203,204],{},"TCO per 1M Tokens",[68,206,207],{},"$1.48",[68,209,210],{},"$1.95",[68,212,213],{},[32,214,215],{},"-24% (AMD有利)",[25,217,218],{},"📉 これは理論上のTCOモデルによるもので、\nAMDは「同等性能をより低コストで実現可能」と主張しています。",[25,220,221],{},"しかし、多数の前提（電力単価・稼働率・サポート費用など）がこの比較には含まれており、現実的な価格差は20〜30%の範囲にとどまると考えられます。",[37,223],{},[20,225,227],{"id":226},"_32-実測ベースsemianalysis-inferencemax-v1","3.2 実測ベース（SemiAnalysis / InferenceMAX v1）",[44,229,230,243],{},[47,231,232],{},[50,233,234,237,240],{},[53,235,236],{},"条件",[53,238,239],{},"結果",[53,241,242],{},"コメント",[63,244,245,258,269],{},[50,246,247,250,255],{},[68,248,249],{},"同一ワークロード（SGLang + TRT-LLM）",[68,251,252],{},[32,253,254],{},"B200がCost per Million Tokensで優位",[68,256,257],{},"実測スループットが高く、CUDA最適化が効いている",[50,259,260,263,266],{},[68,261,262],{},"消費電力効率",[68,264,265],{},"MI355Xやや有利（TDP比）",[68,267,268],{},"ただし液冷・PUE次第で差は縮小",[50,270,271,274,277],{},[68,272,273],{},"ソフト最適化負荷",[68,275,276],{},"B200が低い",[68,278,279],{},"ROCmは改善中だがエコシステム差が大きい",[25,281,282,283,286],{},"🧮 ",[32,284,285],{},"実運用条件に寄せるほど、B200優位が明確になる","。\nAMDが理論値で示すTCO差は、運用・最適化・スケール効率でほぼ相殺されます。",[37,288],{},[15,290,292],{"id":291},"_4-実効コストを左右する要因","⚡ 4. 実効コストを左右する要因",[25,294,295],{},[32,296,297],{},"コスト構成要素の比較",[299,300,305],"pre",{"className":301,"code":302,"language":303,"meta":304,"style":304},"language-mermaid shiki shiki-themes vitesse-light vitesse-light","graph TB\n    subgraph \"AMD MI355X\"\n        A1[GPU本体価格\u003Cbr/>-30% vs NVIDIA]\n        A2[メモリ容量\u003Cbr/>288GB 優位]\n        A3[消費電力\u003Cbr/>1400W]\n        A4[ROCm最適化\u003Cbr/>追加コスト]\n        A5[8GPU構成\u003Cbr/>Infinity Fabric]\n    end\n\n    subgraph \"NVIDIA B200\"\n        B1[GPU本体価格\u003Cbr/>基準価格]\n        B2[メモリ容量\u003Cbr/>192GB]\n        B3[消費電力\u003Cbr/>1787W]\n        B4[CUDA最適化\u003Cbr/>エコシステム成熟]\n        B5[72GPU構成\u003Cbr/>NVL72]\n    end\n\n    A1 -->|CAPEX| C[総所有コスト]\n    A2 -->|性能| C\n    A3 -->|OPEX| C\n    A4 -->|OPEX| C\n    A5 -->|スケール効率| C\n\n    B1 -->|CAPEX| C\n    B2 -->|性能| C\n    B3 -->|OPEX| C\n    B4 -->|OPEX削減| C\n    B5 -->|スケール効率| C\n\n    style A1 fill:#90EE90\n    style A2 fill:#90EE90\n    style B4 fill:#87CEEB\n    style B5 fill:#87CEEB\n","mermaid","",[306,307,308,316,322,328,334,340,346,352,358,365,371,377,383,389,395,401,406,411,417,423,429,435,441,446,452,458,464,470,476,481,487,493,499],"code",{"__ignoreMap":304},[309,310,313],"span",{"class":311,"line":312},"line",1,[309,314,315],{},"graph TB\n",[309,317,319],{"class":311,"line":318},2,[309,320,321],{},"    subgraph \"AMD MI355X\"\n",[309,323,325],{"class":311,"line":324},3,[309,326,327],{},"        A1[GPU本体価格\u003Cbr/>-30% vs NVIDIA]\n",[309,329,331],{"class":311,"line":330},4,[309,332,333],{},"        A2[メモリ容量\u003Cbr/>288GB 優位]\n",[309,335,337],{"class":311,"line":336},5,[309,338,339],{},"        A3[消費電力\u003Cbr/>1400W]\n",[309,341,343],{"class":311,"line":342},6,[309,344,345],{},"        A4[ROCm最適化\u003Cbr/>追加コスト]\n",[309,347,349],{"class":311,"line":348},7,[309,350,351],{},"        A5[8GPU構成\u003Cbr/>Infinity Fabric]\n",[309,353,355],{"class":311,"line":354},8,[309,356,357],{},"    end\n",[309,359,361],{"class":311,"line":360},9,[309,362,364],{"emptyLinePlaceholder":363},true,"\n",[309,366,368],{"class":311,"line":367},10,[309,369,370],{},"    subgraph \"NVIDIA B200\"\n",[309,372,374],{"class":311,"line":373},11,[309,375,376],{},"        B1[GPU本体価格\u003Cbr/>基準価格]\n",[309,378,380],{"class":311,"line":379},12,[309,381,382],{},"        B2[メモリ容量\u003Cbr/>192GB]\n",[309,384,386],{"class":311,"line":385},13,[309,387,388],{},"        B3[消費電力\u003Cbr/>1787W]\n",[309,390,392],{"class":311,"line":391},14,[309,393,394],{},"        B4[CUDA最適化\u003Cbr/>エコシステム成熟]\n",[309,396,398],{"class":311,"line":397},15,[309,399,400],{},"        B5[72GPU構成\u003Cbr/>NVL72]\n",[309,402,404],{"class":311,"line":403},16,[309,405,357],{},[309,407,409],{"class":311,"line":408},17,[309,410,364],{"emptyLinePlaceholder":363},[309,412,414],{"class":311,"line":413},18,[309,415,416],{},"    A1 -->|CAPEX| C[総所有コスト]\n",[309,418,420],{"class":311,"line":419},19,[309,421,422],{},"    A2 -->|性能| C\n",[309,424,426],{"class":311,"line":425},20,[309,427,428],{},"    A3 -->|OPEX| C\n",[309,430,432],{"class":311,"line":431},21,[309,433,434],{},"    A4 -->|OPEX| C\n",[309,436,438],{"class":311,"line":437},22,[309,439,440],{},"    A5 -->|スケール効率| C\n",[309,442,444],{"class":311,"line":443},23,[309,445,364],{"emptyLinePlaceholder":363},[309,447,449],{"class":311,"line":448},24,[309,450,451],{},"    B1 -->|CAPEX| C\n",[309,453,455],{"class":311,"line":454},25,[309,456,457],{},"    B2 -->|性能| C\n",[309,459,461],{"class":311,"line":460},26,[309,462,463],{},"    B3 -->|OPEX| C\n",[309,465,467],{"class":311,"line":466},27,[309,468,469],{},"    B4 -->|OPEX削減| C\n",[309,471,473],{"class":311,"line":472},28,[309,474,475],{},"    B5 -->|スケール効率| C\n",[309,477,479],{"class":311,"line":478},29,[309,480,364],{"emptyLinePlaceholder":363},[309,482,484],{"class":311,"line":483},30,[309,485,486],{},"    style A1 fill:#90EE90\n",[309,488,490],{"class":311,"line":489},31,[309,491,492],{},"    style A2 fill:#90EE90\n",[309,494,496],{"class":311,"line":495},32,[309,497,498],{},"    style B4 fill:#87CEEB\n",[309,500,502],{"class":311,"line":501},33,[309,503,504],{},"    style B5 fill:#87CEEB\n",[20,506,508],{"id":507},"_1-スケール効率","(1) スケール効率",[149,510,511,517],{},[152,512,513,516],{},[32,514,515],{},"B200 NVL72構成","は72 GPUを一体化可能で、推論・トレーニング両方で効率が高い。",[152,518,519],{},"MI355Xは8 GPU単位のInfinity Fabric構成で、通信帯域面で不利。",[20,521,523],{"id":522},"_2-電力冷却コスト","(2) 電力・冷却コスト",[149,525,526,529],{},[152,527,528],{},"MI355Xは1400 Wと高密度だが、液冷前提。",[152,530,531,532,535],{},"B200はシステム全体で14.3 kW（8GPU）。\n→ データセンターPUE（1.1〜1.4）を考慮すると、OPEX差は",[32,533,534],{},"年率で数％程度","。",[20,537,539],{"id":538},"_3-ソフトウェア最適化","(3) ソフトウェア最適化",[149,541,542,545],{},[152,543,544],{},"NVIDIAのCUDA + TensorRT-LLMが圧倒的に成熟。",[152,546,547],{},"AMDのROCmは対応拡大中だが、最適化工数・サポートコストが潜在的負担に。",[37,549],{},[15,551,553],{"id":552},"_5-総合評価","📊 5. 総合評価",[44,555,556,568],{},[47,557,558],{},[50,559,560,563,566],{},[53,561,562],{},"観点",[53,564,565],{},"優位",[53,567,242],{},[63,569,570,581,592,603,613],{},[50,571,572,575,578],{},[68,573,574],{},"理論TCO",[68,576,577],{},"🟢 AMD",[68,579,580],{},"$1.48 vs $1.95 で24%安価（仮定ベース）",[50,582,583,586,589],{},[68,584,585],{},"実測スループット",[68,587,588],{},"🟢 NVIDIA",[68,590,591],{},"高負荷LLMで20〜30%高性能",[50,593,594,597,600],{},[68,595,596],{},"電力効率",[68,598,599],{},"⚪ AMD（僅差）",[68,601,602],{},"高TDPだがシステム全体では拮抗",[50,604,605,608,610],{},[68,606,607],{},"ソフト・運用性",[68,609,588],{},[68,611,612],{},"エコシステムの安定性と人材層",[50,614,615,618,620],{},[68,616,617],{},"スケーラビリティ",[68,619,588],{},[68,621,622],{},"NVL72構成で圧倒的スケール",[37,624],{},[15,626,628],{"id":627},"_6-結論","🧭 6. 結論",[149,630,631,637,643],{},[152,632,633,634,535],{},"AMD ",[32,635,636],{},"MI355Xは小〜中規模LLM推論で高い価格性能比を発揮",[152,638,639,640,535],{},"しかし",[32,641,642],{},"NVIDIA B200は大規模運用・実測性能・安定性で依然優位",[152,644,645,648],{},[32,646,647],{},"1百万トークンあたりの実効コスト差は±20〜30％の範囲","に収束する見込み。",[25,650,651],{},"要するに、",[653,654,655],"blockquote",{},[25,656,657],{},"「AMDは理論上安い、NVIDIAは実運用で強い」\nというのが現時点の最も現実的な評価です。",[37,659],{},[15,661,663],{"id":662},"参考資料","📚 参考資料",[149,665,666,676,684,692],{},[152,667,668,669],{},"AMD公式仕様書: ",[670,671,675],"a",{"href":672,"rel":673},"https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html",[674],"nofollow","AMD Instinct MI355X",[152,677,678,679],{},"NVIDIA公式: ",[670,680,683],{"href":681,"rel":682},"https://www.nvidia.com/en-us/data-center/dgx-b200/",[674],"DGX B200",[152,685,686,687,691],{},"SemiAnalysis: ",[688,689,690],"em",{},"“InferenceMAX v1 Cost per Million Tokens Benchmark”"," (May 2025)",[152,693,694],{},"Inference & AI Infra Reports, 2025 Q2〜Q3",[37,696],{},[698,699,700],"style",{},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":304,"searchDepth":318,"depth":318,"links":702},[703,706,707,708,712,717,718,719],{"id":17,"depth":318,"text":18,"children":704},[705],{"id":22,"depth":324,"text":23},{"id":41,"depth":318,"text":42},{"id":143,"depth":318,"text":144},{"id":174,"depth":318,"text":175,"children":709},[710,711],{"id":178,"depth":324,"text":179},{"id":226,"depth":324,"text":227},{"id":291,"depth":318,"text":292,"children":713},[714,715,716],{"id":507,"depth":324,"text":508},{"id":522,"depth":324,"text":523},{"id":538,"depth":324,"text":539},{"id":552,"depth":318,"text":553},{"id":627,"depth":318,"text":628},{"id":662,"depth":318,"text":663},null,"AMDのMI355XとNVIDIAのB200を、1百万トークンあたりのコスト効率で徹底比較。理論TCOと実測性能の差を分析","md",{},"/amd-mi355x-vs-nvidia-b200-cost-analysis",false,"2025-10-17T00:00:00.000Z",{"title":5,"description":721},"2025-10/2025-10-17/amd-mi355x-vs-nvidia-b200-cost-analysis",[730,731,732,733,734,735],"AMD","NVIDIA","AI","GPU","cost-analysis","memo","rhPVLPvsKBs5GfgODcHlWZHrvP_SLMrf-4lOfgfZ1Cg",[],[739,743,747,751,755],{"title":740,"path":741,"publishedAt":742},"CUDA Programming Guide Part 1を小学生にもわかるように読む","/cuda-programming-child-friendly-guide","2026-06-05T00:00:00.000Z",{"title":744,"path":745,"publishedAt":746},"Vera Rubin量産移行、Agentic AIで50倍成長へ","/nvidia_ces_qa_ja","2026-01-07T00:00:00.000Z",{"title":748,"path":749,"publishedAt":750},"Nvidiaの堀：競合他社が直面する「無理ゲー」な9つの壁","/nvidia-moat-impossible-game","2025-11-24T00:00:00.000Z",{"title":752,"path":753,"publishedAt":754},"AMDはサーバーCPUの市場規模予想を2.6か月で1200億→2200億ドルに書き換えた｜土台の260億ドルは動かさず、成長率だけを18%→35%→50%超と3回上げている","/amd-server-cpu-tam-120b-to-220b","2026-07-30T00:00:00.000Z",{"title":756,"path":757,"publishedAt":754},"AI半導体のTAMは誰がどう出しているか｜NVIDIAは年3〜4兆ドル、Samsungは金額を出さない。同じ「TAM」の一語で3つの別物が語られている","/who-publishes-ai-tam-and-who-does-not","https://log.eurekapu.com/og/blog/amd-mi355x-vs-nvidia-b200-cost-analysis.png?v=2025-10-17T00%3A00%3A00.000Z&title=AMD%20MI355X%20vs%20NVIDIA%20B200%20-%201%E7%99%BE%E4%B8%87%E3%83%88%E3%83%BC%E3%82%AF%E3%83%B3%E3%81%82%E3%81%9F%E3%82%8A%E3%82%B3%E3%82%B9%E3%83%88%E3%81%A7%E8%A6%8B%E3%82%8B%E7%9C%9F%E3%81%AE%E5%8B%9D%E8%80%85&author=Kei%20Komatsu&sig=2c00e75e12d5d5fc",1785654656995]