{
 "generated": "2026-11-10",
 "sha": "c82ffc020d8a",
 "node": {
  "id": "event:sig-0121",
  "type": "Event",
  "name": "Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍",
  "slug": "sig-0121",
  "kind": "signal",
  "kind_zh": "行业信号",
  "date": "2026-11-10",
  "summary": "11-09 发布的报告。Epoch 认为用 FLOP/s 衡量推理容量会严重低估真实能力——对推理、尤其是现在的 agentic 负载来说，显存容量与带宽、芯片间互连的带宽与延迟往往比算术吞吐更要紧，于是另建了一套 Inference Performance Index，用四个代表性模型、在每请求 50-400 token/s 的输出速度区间内衡量容量，并显式处理「同时服务多少并发请求」与「单请求多快」之间的权衡。以 GB300 NVL72 为基准 1.0 GB300e：Vera Rubin（VR200）2.18、H200 HGX 0.12；最关键的一条是 GB300 在 agentic 任务上的推理容量约为 H100 的 25 倍，而两者 FLOP/s 只差 2.5 倍。意味着按算力账本做的产能推算，可能系统性低估了这一代硬件的推理供给。",
  "url": "https://epoch.ai/publications/inference-performance-index",
  "evidence": [
   {
    "kind": "own",
    "label": "本站页面",
    "url": "https://xiaopingfeng.com/graph/signals/#sig-0121",
    "text": "AI 行业信号流（独立研究机构）：Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍"
   },
   {
    "kind": "media",
    "url": "https://epoch.ai/publications/inference-performance-index",
    "text": "信号的原文出处（Epoch AI）"
   }
  ],
  "entry": "derived",
  "category": "research",
  "category_zh": "独立研究机构",
  "source_name": "Epoch AI",
  "source_id": "org-epochai",
  "links": [
   {
    "label": "Epoch AI",
    "url": "https://epoch.ai/publications/inference-performance-index"
   }
  ],
  "degree": 0
 },
 "page": "https://xiaopingfeng.com/graph/e/event/sig-0121/",
 "out": [
  {
   "type": "placed_in",
   "sentence": "Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍出现在 Landscape 的「独立研究机构」",
   "direction": "out",
   "id": "layer:ai-industry-signals--5",
   "name": "独立研究机构",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0121",
     "text": "AI 行业信号流（独立研究机构）：Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍"
    }
   ]
  },
  {
   "type": "involves",
   "sentence": "Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍涉及Epoch AI",
   "direction": "out",
   "id": "org:epoch-ai",
   "name": "Epoch AI",
   "role": "issuer",
   "note": "信源：Epoch AI",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "url": "https://xiaopingfeng.com/graph/signals/#sig-0121",
     "text": "AI 行业信号流（独立研究机构）：Epoch AI 推出推理性能指数 IPI：按 agentic 负载算，GB300 的推理容量是 H100 的约 25 倍，而 FLOP/s 只有 2.5 倍"
    }
   ]
  }
 ],
 "in": []
}