Whisper.cpp Large V3 Turbo (Concurrency)

Scoring

Primary metric: real_time_factor · Tie-breakers: throughput, gpu_pct

Metrics: real_time_factor, throughput, cpu_pct, ram_mb, gpu_pct, energy_wh, temperature_c

Changelog

• v1.0.0 — New profile for whisper-large-v3-turbo. Mirrors its whisper-large-v3 sibling field-for-field (turbo is the same encoder with a pruned 4-layer decoder -- no config-surface change) -- added per external feedback (2026-08-05 Babbl letter). Verified real support first, not assumed -- whisper.cpp's official ggml-large-v3-turbo.bin (download-ggml-model.sh large-v3-turbo) and faster-whisper's own _MODELS mapping ("large-v3-turbo"/"turbo" both resolve to mobiuslabsgmbh/faster-whisper-large-v3-turbo, confirmed by reading faster_whisper/utils.py directly).

Full definition

{
  "id": "whispercpp-large-v3-turbo-concurrency",
  "version": "1.0.0",
  "title": "Whisper.cpp Large V3 Turbo (Concurrency)",
  "benchmark_type": "concurrency",
  "runtime": {
    "name": "whisper-cpp",
    "min_version": "1.5.0"
  },
  "model": {
    "name": "whisper-large-v3-turbo",
    "beam_size": 5,
    "temperature": 0
  },
  "configuration": {
    "threads": 4,
    "concurrency": 1,
    "duration_s": 30
  },
  "scoring": {
    "primary_metric": "real_time_factor",
    "tie_breakers": [
      "throughput",
      "gpu_pct"
    ]
  },
  "metrics": [
    "real_time_factor",
    "throughput",
    "cpu_pct",
    "ram_mb",
    "gpu_pct",
    "energy_wh",
    "temperature_c"
  ],
  "overridable": {
    "concurrency": {
      "range": {
        "min": 1,
        "max": 16
      }
    },
    "threads": {
      "range": {
        "min": 1,
        "max": 16
      }
    }
  },
  "changelog": [
    {
      "version": "1.0.0",
      "notes": "New profile for whisper-large-v3-turbo. Mirrors its whisper-large-v3 sibling field-for-field (turbo is the same encoder with a pruned 4-layer decoder -- no config-surface change) -- added per external feedback (2026-08-05 Babbl letter). Verified real support first, not assumed -- whisper.cpp's official ggml-large-v3-turbo.bin (download-ggml-model.sh large-v3-turbo) and faster-whisper's own _MODELS mapping (\"large-v3-turbo\"/\"turbo\" both resolve to mobiuslabsgmbh/faster-whisper-large-v3-turbo, confirmed by reading faster_whisper/utils.py directly)."
    }
  ]
}