{
  "title": "LLMs aren't flat. They're split.",
  "footnote": "Decode stays low across the board, while prefill spans from nothing to 54%. The multi-card decode points sit inside their own run-to-run spread.",
  "height": 420,
  "kind": "scatter",
  "max": 10,
  "min": -2,
  "options": {
    "xMin": -2,
    "xMax": 70,
    "pointLabelKey": "show",
    "xSuffix": "%"
  },
  "rows": [
    {
      "label": "Qwen3.5-9B, Strix-B",
      "show": "Qwen3.5-9B, Strix-B",
      "x": 54,
      "y": 3.9,
      "_group": "BF16 on vLLM",
      "_color": 1,
      "_pos": "top",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Qwen3.5-9B, Zotac",
      "show": "",
      "x": 47.5,
      "y": 4.6,
      "_group": "BF16 on vLLM",
      "_color": 1,
      "_pos": "top",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Qwen3.5-9B, PNY",
      "show": "",
      "x": 60.4,
      "y": 5.1,
      "_group": "BF16 on vLLM",
      "_color": 1,
      "_pos": "top",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Gemma 4 E4B",
      "show": "Gemma 4 E4B",
      "x": 38.1,
      "y": 4.5,
      "_group": "BF16 on vLLM",
      "_color": 1,
      "_pos": "top",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Qwen3-VL-8B (INT4)",
      "show": "Qwen3-VL-8B (INT4)",
      "x": 52.3,
      "y": 5.4,
      "_group": "INT4 on vLLM",
      "_color": 2,
      "_pos": "top",
      "_size": 12,
      "_phone": true
    },
    {
      "label": "Ornith-1.5-9B",
      "show": "Ornith-1.5-9B",
      "x": 34.8,
      "y": 0,
      "_group": "BF16 GGUF",
      "_color": 3,
      "_pos": "top",
      "_size": 12,
      "_phone": true
    },
    {
      "label": "Gemma 4 12B, 2 cards",
      "show": "Gemma 4 12B, 2 cards",
      "x": 17.3,
      "y": 3.7,
      "_group": "Split across cards",
      "_color": 4,
      "_pos": "top",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Qwen3.8-27B, 4 cards",
      "show": "Qwen3.8-27B, 4 cards",
      "x": 10.8,
      "y": 1.1,
      "_group": "Split across cards",
      "_color": 4,
      "_pos": "bottom",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Ornith-1.5-9B Q4",
      "show": "",
      "x": 1.8,
      "y": 0.6,
      "_group": "Q4 GGUF",
      "_color": "muted",
      "_pos": "bottom",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Ornith-1.5-35B Q4",
      "show": "",
      "x": 0.7,
      "y": -0.4,
      "_group": "Q4 GGUF",
      "_color": "muted",
      "_pos": "bottom",
      "_size": 12,
      "_phone": false
    },
    {
      "label": "Qwen3.8-27B Q4",
      "show": "Qwen3.8-27B Q4",
      "x": 0.3,
      "y": -0.3,
      "_group": "Q4 GGUF",
      "_color": "muted",
      "_pos": "bottom",
      "_size": 12,
      "_phone": true
    },
    {
      "label": "Gemma 12B QAT Q4",
      "show": "",
      "x": 1.3,
      "y": 0.3,
      "_group": "Q4 GGUF",
      "_color": "muted",
      "_pos": "bottom",
      "_size": 12,
      "_phone": false
    }
  ],
  "series": [
    {
      "key": "y",
      "label": "Decode gain",
      "suffix": "%",
      "decimals": 1
    }
  ],
  "subtitle": "Every LLM setup at an 8K prompt. Across: how much prefill gained. Up: how much decode gained.",
  "x": "x",
  "xLabel": "Prefill gain (%)",
  "yLabel": "Decode gain (%)",
  "__hash__": "5NtFbbaRE29KHIRUaRvhCZBp4tFTG-EA5U8_6xPQLp8"
}