{
  "schema": "anvil-serving.benchmark-graph-manifest/v1",
  "title": "RTX5090: short-output configuration comparison",
  "subtitle": "Different private checkpoints/runtimes;120 valid requests each;16 words;C1;not a model ranking.",
  "output": "capacity-comparison.svg",
  "data_output": "capacity-graph-data.json",
  "metric_semantics": {
    "context": "4096 planned; actual usage3609-3687",
    "reasoning": "disabled",
    "cache": "unique request prefixes with canaries",
    "output": "16 exact words; native completion counts28 for Ignis and29 for both NInfer cells.",
    "p99": "Nearest-rank empirical tail estimate; not an SLO",
    "identity": "Distinct private checkpoints/runtime profiles; reference and QAT configuredC1; Ignis configured4lanes but issuedC1. Same configured32K limit and16-word target; actual input/output tokenization differs. No isolated causal claim."
  },
  "charts": [
    {
      "title": "First-token median",
      "metric": "metrics.ttft_p50_ms",
      "x_label": "Actual input3609-3687 tokens;16 words;28-29 completion tokens",
      "y_label": "milliseconds",
      "series": [
        {
          "label": "Reference32K / DF7",
          "points": [
            {
              "artifact": "raw/reference-matched-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Candidate2 / no-spec32K",
          "points": [
            {
              "artifact": "raw/qat-reverse-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Compact32K /4L",
          "points": [
            {
              "artifact": "raw/ignis-budget4096-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        }
      ]
    },
    {
      "title": "End-to-end median",
      "metric": "metrics.e2e_p50_ms",
      "x_label": "Actual input3609-3687 tokens;16 words;28-29 completion tokens",
      "y_label": "milliseconds",
      "series": [
        {
          "label": "Reference32K / DF7",
          "points": [
            {
              "artifact": "raw/reference-matched-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Candidate2 / no-spec32K",
          "points": [
            {
              "artifact": "raw/qat-reverse-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Compact32K /4L",
          "points": [
            {
              "artifact": "raw/ignis-budget4096-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        }
      ]
    },
    {
      "title": "Decode median",
      "metric": "metrics.decode_tok_s_p50",
      "x_label": "Actual input3609-3687 tokens;16 words;28-29 completion tokens",
      "y_label": "tokens/second",
      "series": [
        {
          "label": "Reference32K / DF7",
          "points": [
            {
              "artifact": "raw/reference-matched-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Candidate2 / no-spec32K",
          "points": [
            {
              "artifact": "raw/qat-reverse-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Compact32K /4L",
          "points": [
            {
              "artifact": "raw/ignis-budget4096-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        }
      ]
    },
    {
      "title": "First-token p99",
      "metric": "metrics.ttft_p99_ms",
      "x_label": "Actual input3609-3687 tokens;16 words;28-29 completion tokens",
      "y_label": "milliseconds;120 samples;nearest rank",
      "series": [
        {
          "label": "Reference32K / DF7",
          "points": [
            {
              "artifact": "raw/reference-matched-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Candidate2 / no-spec32K",
          "points": [
            {
              "artifact": "raw/qat-reverse-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        },
        {
          "label": "Compact32K /4L",
          "points": [
            {
              "artifact": "raw/ignis-budget4096-capacity16-120.json",
              "x": "C1 / unique"
            }
          ]
        }
      ]
    }
  ]
}