{
  "benchmark": "multi_gpu",
  "timestamp": "2026-03-19T12:21:06.000529+00:00",
  "n_gpus_available": 4,
  "gpu_names": [
    "NVIDIA L4",
    "NVIDIA L4",
    "NVIDIA L4",
    "NVIDIA L4"
  ],
  "inference": {
    "gpu_1": {
      "n_gpus": 1,
      "batch_results": {
        "batch_1": {
          "p50_ms": 128.19,
          "p95_ms": 130.86,
          "mean_ms": 128.42,
          "fps": 7.8,
          "per_sample_ms": 128.42
        },
        "batch_4": {
          "p50_ms": 430.18,
          "p95_ms": 440.71,
          "mean_ms": 429.76,
          "fps": 9.3,
          "per_sample_ms": 107.44
        },
        "batch_8": {
          "p50_ms": 853.99,
          "p95_ms": 873.4,
          "mean_ms": 857.72,
          "fps": 9.3,
          "per_sample_ms": 107.21
        },
        "batch_16": {
          "p50_ms": 1734.01,
          "p95_ms": 1759.55,
          "mean_ms": 1727.97,
          "fps": 9.3,
          "per_sample_ms": 108.0
        }
      },
      "memory": {
        "gpu_0_allocated_gb": 3.65,
        "gpu_0_reserved_gb": 5.07
      }
    },
    "gpu_2": {
      "n_gpus": 2,
      "batch_results": {
        "batch_1": {
          "p50_ms": 162.23,
          "p95_ms": 168.32,
          "mean_ms": 164.65,
          "fps": 6.1,
          "per_sample_ms": 164.65
        },
        "batch_4": {
          "p50_ms": 611.69,
          "p95_ms": 613.66,
          "mean_ms": 611.27,
          "fps": 6.5,
          "per_sample_ms": 152.82
        },
        "batch_8": {
          "p50_ms": 799.07,
          "p95_ms": 802.8,
          "mean_ms": 799.2,
          "fps": 10.0,
          "per_sample_ms": 99.9
        },
        "batch_16": {
          "p50_ms": 1185.3,
          "p95_ms": 1190.2,
          "mean_ms": 1184.72,
          "fps": 13.5,
          "per_sample_ms": 74.04
        }
      },
      "memory": {
        "gpu_0_allocated_gb": 3.66,
        "gpu_0_reserved_gb": 4.47,
        "gpu_1_allocated_gb": 0.01,
        "gpu_1_reserved_gb": 4.47
      }
    },
    "gpu_4": {
      "n_gpus": 4,
      "batch_results": {
        "batch_1": {
          "p50_ms": 164.7,
          "p95_ms": 171.49,
          "mean_ms": 167.37,
          "fps": 6.0,
          "per_sample_ms": 167.37
        },
        "batch_4": {
          "p50_ms": 913.4,
          "p95_ms": 915.7,
          "mean_ms": 912.74,
          "fps": 4.4,
          "per_sample_ms": 228.19
        },
        "batch_8": {
          "p50_ms": 1003.53,
          "p95_ms": 1007.34,
          "mean_ms": 1002.41,
          "fps": 8.0,
          "per_sample_ms": 125.3
        },
        "batch_16": {
          "p50_ms": 1178.17,
          "p95_ms": 1182.78,
          "mean_ms": 1178.2,
          "fps": 13.6,
          "per_sample_ms": 73.64
        }
      },
      "memory": {
        "gpu_0_allocated_gb": 3.66,
        "gpu_0_reserved_gb": 4.49,
        "gpu_1_allocated_gb": 0.01,
        "gpu_1_reserved_gb": 4.16,
        "gpu_2_allocated_gb": 0.01,
        "gpu_2_reserved_gb": 4.16,
        "gpu_3_allocated_gb": 0.01,
        "gpu_3_reserved_gb": 4.16
      }
    }
  },
  "training": {
    "gpu_1": {
      "n_gpus": 1,
      "batch_size": 2,
      "n_steps": 30,
      "step_time_mean_ms": 432.4,
      "steps_per_sec": 2.31,
      "samples_per_sec": 4.63,
      "loss_start": 4.0196,
      "loss_end": 1.7553,
      "loss_reduction_pct": 56.3,
      "memory": {
        "gpu_0_peak_gb": 9.0
      }
    },
    "gpu_2": {
      "n_gpus": 2,
      "batch_size": 4,
      "n_steps": 30,
      "step_time_mean_ms": 1264.8,
      "steps_per_sec": 0.79,
      "samples_per_sec": 3.16,
      "loss_start": 1.3165,
      "loss_end": 1.4857,
      "loss_reduction_pct": -12.9,
      "memory": {
        "gpu_0_peak_gb": 14.59,
        "gpu_1_peak_gb": 4.07
      }
    },
    "gpu_4": {
      "n_gpus": 4,
      "batch_size": 8,
      "n_steps": 30,
      "step_time_mean_ms": 2005.1,
      "steps_per_sec": 0.5,
      "samples_per_sec": 3.99,
      "loss_start": 6.7918,
      "loss_end": 1.182,
      "loss_reduction_pct": 82.6,
      "memory": {
        "gpu_0_peak_gb": 14.6,
        "gpu_1_peak_gb": 4.07,
        "gpu_2_peak_gb": 4.07,
        "gpu_3_peak_gb": 4.07
      }
    }
  },
  "fp16": {
    "fp16_gpu_1": {
      "n_gpus": 1,
      "precision": "fp16",
      "batch_results": {
        "batch_4": {
          "p50_ms": 122.14,
          "fps": 32.7,
          "per_sample_ms": 30.56
        },
        "batch_8": {
          "p50_ms": 234.11,
          "fps": 34.2,
          "per_sample_ms": 29.26
        },
        "batch_16": {
          "p50_ms": 486.22,
          "fps": 32.9,
          "per_sample_ms": 30.43
        },
        "batch_32": {
          "p50_ms": 950.24,
          "fps": 33.6,
          "per_sample_ms": 29.73
        }
      }
    },
    "fp16_gpu_4": {
      "n_gpus": 4,
      "precision": "fp16",
      "batch_results": {
        "batch_4": {
          "p50_ms": 901.49,
          "fps": 4.4,
          "per_sample_ms": 225.52
        },
        "batch_8": {
          "p50_ms": 903.67,
          "fps": 8.8,
          "per_sample_ms": 113.02
        },
        "batch_16": {
          "p50_ms": 911.97,
          "fps": 17.5,
          "per_sample_ms": 57.07
        },
        "batch_32": {
          "p50_ms": 1013.11,
          "fps": 31.6,
          "per_sample_ms": 31.67
        }
      }
    }
  }
}