{
  "title": "One big card beats two, and the gap grows with load",
  "footnote": "Ornith-1.5-35B-A3B int4 on vLLM. The horizontal spacing is by test, so it isn't to scale.",
  "height": 340,
  "kind": "area",
  "max": 2300,
  "options": {
    "xTicks": [
      "1",
      "8",
      "32"
    ],
    "points": [
      {
        "x": 0,
        "y": 200,
        "text": "200",
        "sub": "vs 153.9",
        "align": "right"
      },
      {
        "x": 1,
        "y": 1049.2,
        "text": "1,049",
        "sub": "vs 667.6",
        "align": "center"
      },
      {
        "x": 2,
        "y": 2072.3,
        "text": "2,072",
        "sub": "vs 758.4",
        "align": "left"
      }
    ]
  },
  "rows": [
    {
      "i": 0,
      "label": "1",
      "pair": 153.9,
      "gap": 46.1,
      "single": 200
    },
    {
      "i": 1,
      "label": "8",
      "pair": 667.6,
      "gap": 381.6,
      "single": 1049.2
    },
    {
      "i": 2,
      "label": "32",
      "pair": 758.4,
      "gap": 1313.9,
      "single": 2072.3
    }
  ],
  "series": [
    {
      "key": "pair",
      "label": "2x RTX 3090, tensor-parallel",
      "color": "muted",
      "decimals": 1
    },
    {
      "key": "gap",
      "label": "What 1x RTX PRO 6000 adds",
      "color": 1,
      "decimals": 1
    }
  ],
  "stats": [
    {
      "value": "2.7x",
      "label": "throughput at 32 concurrent requests"
    },
    {
      "value": "11.76s → 1.57s",
      "label": "median time-to-first-token at 32 concurrent"
    }
  ],
  "subtitle": "Aggregate tokens/sec, same model and quant on both setups. The green wedge is what the single card adds on top of the pair.",
  "x": "i",
  "xLabel": "Concurrent requests",
  "yLabel": "Tokens/sec",
  "__hash__": "T_CvwGOBFxVQ7ELJ31wX_Oj_DHCa2ZLucov0Zrmslmo"
}