{
  "chart": "s2s-chart/1",
  "kind": "bars",
  "title": "What does 128 GB of eight-channel memory buy? A 180B mixture from memory alone, at three thread counts",
  "subtitle": "One 180-billion-parameter mixture, Qwen3.8-Flash-Next at IQ3_S (about three bits a weight; about 6 billion parameters awake per token), from 128 GB of memory alone beside that machine's live services, at three thread counts: writing, prompt reading, the first word and the energy per request (NOT WALL).",
  "series": [
    "Measured"
  ],
  "panels": [
    {
      "heading": "Writing speed, tok/s (client clock)",
      "unit": "tok/s",
      "rows": [
        {
          "label": "8 threads",
          "notes": "",
          "values": [
            {
              "value": "6.51",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[0].decode_median 6.51"
            }
          ]
        },
        {
          "label": "16 threads",
          "notes": "",
          "values": [
            {
              "value": "7.74",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[2].decode_median 7.741"
            }
          ]
        },
        {
          "label": "32 threads",
          "notes": "",
          "values": [
            {
              "value": "8.33",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[1].decode_median 8.325"
            }
          ]
        }
      ]
    },
    {
      "heading": "Prompt reading, tok/s (4,038-token prompts)",
      "unit": "tok/s",
      "rows": [
        {
          "label": "8 threads",
          "notes": "",
          "values": [
            {
              "value": "19.6",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[0].prefill_median 19.633"
            }
          ]
        },
        {
          "label": "16 threads",
          "notes": "",
          "values": [
            {
              "value": "33.0",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[2].prefill_median 33.027"
            }
          ]
        },
        {
          "label": "32 threads",
          "notes": "",
          "values": [
            {
              "value": "41.1",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[1].prefill_median 41.069"
            }
          ]
        }
      ]
    },
    {
      "heading": "First word, s",
      "unit": "s",
      "rows": [
        {
          "label": "8 threads",
          "notes": "",
          "values": [
            {
              "value": "205.70",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[0].first_word_median_ms 205698.7 ÷ 1,000"
            }
          ]
        },
        {
          "label": "16 threads",
          "notes": "",
          "values": [
            {
              "value": "122.28",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[2].first_word_median_ms 122276.4 ÷ 1,000"
            }
          ]
        },
        {
          "label": "32 threads",
          "notes": "",
          "values": [
            {
              "value": "98.35",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[1].first_word_median_ms 98346.5 ÷ 1,000"
            }
          ]
        }
      ]
    },
    {
      "heading": "Energy a request, kJ (NOT WALL, model loaded)",
      "unit": "kJ",
      "rows": [
        {
          "label": "8 threads",
          "notes": "",
          "values": [
            {
              "value": "26.633",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[0].request_kj_median 26.633"
            }
          ]
        },
        {
          "label": "16 threads",
          "notes": "",
          "values": [
            {
              "value": "20.925",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[2].request_kj_median 20.925"
            }
          ]
        },
        {
          "label": "32 threads",
          "notes": "",
          "values": [
            {
              "value": "19.308",
              "source": "/speeds-from-ram-alone/data/eight-channel-desktop/summary.json arms[1].request_kj_median 19.308"
            }
          ]
        }
      ]
    }
  ],
  "source": "Exhibit 105's kit, eight-channel-desktop/summary.json, arms[0] (8 threads), arms[2] (16) and arms[1] (32): decode_median, prefill_median, first_word_median_ms (ms ÷ 1,000) and request_kj_median; llama.cpp build 11514, 2026-11-03. Table T7a and T7b."
}