{
  "kind": "stacked-memory-bars",
  "title": "Persistent model state for one billion parameters and four workers",
  "description": "Stack parameters, gradients, optimizer/master weights at stages 0–3; totals 16, 7, 5.5, 4 decimal GB per worker. Explicit caption: excludes activations, communication buffers, workspaces, and transient gathers.",
  "data": {
    "unit": "decimal GB per worker",
    "stages": [
      "unsharded",
      "stage 1",
      "stage 2",
      "stage 3"
    ],
    "series": [
      {
        "name": "parameters",
        "values": [
          2,
          2,
          2,
          0.5
        ]
      },
      {
        "name": "gradients",
        "values": [
          2,
          2,
          0.5,
          0.5
        ]
      },
      {
        "name": "optimizer and master weights",
        "values": [
          12,
          3,
          3,
          3
        ]
      }
    ],
    "parameters": 1000000000,
    "workers": 4,
    "excluded": [
      "activations",
      "communication buffers",
      "temporary gathers",
      "operator workspaces"
    ]
  },
  "evidence": "Original educational calculation; no measured paper or hardware benchmark."
}