{
  "analytic": "quantization-bits",
  "kind": "declared_analytical_or_constructed_example",
  "experimental_model_result": false,
  "panels": [
    {
      "kind": "bar",
      "labels": [
        "2",
        "3",
        "4",
        "8",
        "16"
      ],
      "series": [
        {
          "label": "GB (10^9 bytes)",
          "y": [
            1.75,
            2.625,
            3.5,
            7.0,
            14.0
          ]
        },
        {
          "label": "GiB (2^30 bytes)",
          "y": [
            1.6298145055770874,
            2.444721758365631,
            3.259629011154175,
            6.51925802230835,
            13.0385160446167
          ]
        }
      ],
      "ylabel": "Raw weight storage"
    }
  ],
  "title": "Raw weight storage is a lower-level accounting quantity",
  "assumptions": "7 x 10^9 scalar weights; packed exactly at labeled bit widths. No scales, zeros, floating layers, KV, activation or workspace.",
  "worked": "7×10^9个权重按4位理想打包为3.5 GB，约3.260 GiB。分组尺度、未量化层、KV及工作区尚未计入，不能称为完整部署显存。",
  "source_ids": [
    "arxiv-2210.17323",
    "arxiv-2306.00978",
    "arxiv-2211.10438"
  ],
  "formula_source": "Article equations and this versioned script",
  "diagram_nodes": [
    "浮点权重与独立校准集",
    "分组 / 尺度 / 异常值",
    "量化打包与元数据",
    "实际内核及反量化",
    "层误差与任务检验",
    "完整内存 / 时延账本"
  ]
}