{
  "id": "supertonic-gpu-first-audio",
  "story": "supertonic-triton-production",
  "title": "FP32 GPU execution reduced first-audio delay",
  "unit": "median seconds to first PCM byte",
  "rows": [
    {
      "label": "English · CPU",
      "value": 1.154,
      "highlight": false
    },
    {
      "label": "English · Triton GPU",
      "value": 0.349,
      "highlight": true
    },
    {
      "label": "Spanish · CPU",
      "value": 1.079,
      "highlight": false
    },
    {
      "label": "Spanish · Triton GPU",
      "value": 0.341,
      "highlight": true
    }
  ],
  "scope": "96 renders · 12 phrases per language · two repeats · paired CPU/GPU",
  "caveat": "NVIDIA Triton ONNX backend with CUDA execution. Excludes ASR, phone capture and call network delay.",
  "kind": "bar",
  "decimals": 3,
  "revision": "7e6fdbe3b1d1aa360346255bfc62945dba694399",
  "sources": [
    {
      "path": "tts/supertonic3-precision-triton-2026-10-13/triton/REPORT.md",
      "url": "https://github.com/groxaxo/experimentos/blob/7e6fdbe3b1d1aa360346255bfc62945dba694399/tts/supertonic3-precision-triton-2026-10-13/triton/REPORT.md",
      "sha256": "018645cc0da78c42af724f74dc9206af34a82bb77c91fd79aac00132d3a63358"
    }
  ],
  "origin": "reviewed report table",
  "maximum": null,
  "threshold": null
}