{
  "batch_e2e": {
    "fp16_baseline": {
      "e2e_s": 7.66,
      "decode_ms_per_tok": 44.1,
      "tokens": "161/161"
    },
    "int8": {
      "e2e_s": 6.24,
      "decode_ms_per_tok": 37.0,
      "tokens": "161/161"
    },
    "int4": {
      "e2e_s": 5.44,
      "decode_ms_per_tok": 32.1,
      "tokens": "161/161"
    },
    "int4_qkv": {
      "e2e_s": 5.26,
      "decode_ms_per_tok": 30.9,
      "tokens": "161/161"
    },
    "int4_qkv_sdpa": {
      "e2e_s": 4.74,
      "decode_ms_per_tok": 28.1,
      "tokens": "161/161"
    },
    "int4_qkv_sdpa_batched_first_chunk": {
      "e2e_s": 4.45,
      "decode_ms_per_tok": 26.6,
      "tokens": "161/161"
    },
    "fp16_qkv_sdpa": {
      "e2e_s": 7.93,
      "decode_ms_per_tok": 45.3,
      "tokens": "161/161"
    },
    "int8_qkv_sdpa": {
      "e2e_s": 6.11,
      "decode_ms_per_tok": 35.6,
      "tokens": "161/161"
    }
  },
  "streaming": {
    "serialized_ttft_s": 1.53,
    "batched_ttft_s": 0.148,
    "chunk_decode_ms": 59.6,
    "tokens": "128/128",
    "passes": 5
  },
  "protocol": {
    "fixture": "12.02s synthetic say audio, 16kHz mono float32",
    "warmup": "1 untimed pass + 3 timed passes (batch), 1 warmup + 5 timed passes (streaming)",
    "hardware": "Apple M4 Max, 64GB unified memory",
    "os": "macOS 27.0 (Darwin 27.0.0)",
    "runtime": "Core AI 1.0.0b2 (beta 4)",
    "backend": "GPU (xctrace: 23,470 GPU Compute intervals, 0 ANE intervals)"
  }
}