{
  "schema_version": 1,
  "slug": "nemotron-asr-streaming-0-6b",
  "model": "nvidia/nemotron-3.5-asr-streaming-0.6b",
  "developer": "NVIDIA",
  "modality": {
    "input": ["audio"],
    "output": ["text"]
  },
  "serving_profile": {
    "engine": ".wave",
    "kernel": "Wave Persistent Kernel (WPK)",
    "hardware": "1× NVIDIA H100 SXM 80 GB",
    "model_step_ms": 80,
    "steps_per_beat": 1,
    "deadline_ms": 80,
    "chunk_setting": "80 ms, the model card's lowest-latency streaming setting",
    "timing_boundary": "benchmark client on the serving host, over loopback, through the gateway ticket and the data plane WebSocket",
    "precision": {
      "encoder": "BF16 with FP32 accumulation",
      "decoder": "FP32",
      "note": "the precision the production NIM runs"
    }
  },
  "headline_result": {
    "status": "measured",
    "streams": 4800,
    "highest_qualified": true,
    "runs": 3,
    "frames_completed_per_run": "7,800,000 of 7,800,000",
    "streams_failed": 0,
    "completion_gap_p99_ms": {
      "min": 94.3,
      "max": 96.6,
      "bound_ms": 120
    },
    "worst_stream_p99_ms": {
      "min": 105.6,
      "max": 106.3
    },
    "causal_latency_p50_ms": {
      "min": 207,
      "max": 211,
      "jitter_hold_periods": 1
    },
    "word_error_rate_vs_fp32_reference_percent": {
      "mean": 0.23,
      "worst_program": 1.34
    }
  },
  "landing_comparisons": {
    "reference": "NVIDIA model card for nemotron-3.5-asr-streaming-0.6b: 240 concurrent streams at the 80 ms setting on one H100, measurement method not stated",
    "streams_per_gpu_multiple": 20,
    "gpu_cost_per_stream_reduction_percent": 95,
    "gpu_cost_assumption": "Equal H100-hour price and utilization; cost per stream is inversely proportional to real-time density.",
    "comparison_note": ".wave at 4,800 streams is a caller-visible measurement under a preregistered protocol; the 240 is a published figure whose method the model card does not state. This is not a matched measured run of the NIM."
  },
  "measured_points": [
    {
      "engine": ".wave, caller-visible",
      "streams": 4800,
      "frames_completed": "7,800,000 / 7,800,000",
      "completion_gap_p99_ms": "94.3–96.6",
      "worst_stream_p99_ms": "105.6–106.3",
      "causal_p50_ms": "207–211",
      "runs": 3
    },
    {
      "engine": ".wave, caller-visible",
      "streams": 5120,
      "frames_completed": "all",
      "completion_gap_p99_ms": "bound held per stream",
      "causal_p50_ms": "not qualified",
      "runs": 1
    },
    {
      "engine": "NVIDIA model card",
      "streams": 240,
      "frames_completed": "not stated",
      "completion_gap_p99_ms": "not stated",
      "causal_p50_ms": "not stated",
      "runs": "not stated"
    },
    {
      "engine": "NeMo 3.0.0 cache-aware streaming script, lockstep batches",
      "streams": 256,
      "frames_completed": "throughput point",
      "completion_gap_p99_ms": "chunk time p99 under 80 ms, no per-stream timing",
      "causal_p50_ms": "not measured",
      "runs": 1
    }
  ],
  "workload": {
    "programs": "23 programs of 130 s built from LibriSpeech utterances with 480 ms of silence between them, one drawn at random per stream",
    "admission": "every stream admitted before any audio flows; all 4,800 on speech together for 73 s in the middle of each run",
    "transcript_rule": "word error rate against the fp32 reference as streamed, at the precision the production NIM runs; token-exact on 12 of 23 programs, text-exact on 15, every difference at a segment boundary on silence"
  },
  "evidence": {
    "article": "../../technical-notes/wpk-persistent-kernel/",
    "model_source": "https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b",
    "study_period": "September 2026"
  }
}