{
  "n_samples": 800,
  "wall_time_s": 2.8710978660019464,
  "throughput_samples_per_s": 278.6390563251729,
  "gpu_total_kernel_ms": 13.113408036530018,
  "gpu_ms_per_sample": 0.01639176004566252,
  "gpu_avg_ms_per_batch": 0.26226816073060033,
  "gpu_peak_mem_MB": 36.26806640625,
  "telemetry_pre": {
    "cpu_freq_current_MHz": 1036.8826964285715,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 4.2,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 0,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1621.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 201.132,
    "gpu_temp_C": 42
  },
  "telemetry_post": {
    "cpu_freq_current_MHz": 1191.0390178571429,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 2.9,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 1,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1641.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 203.233,
    "gpu_temp_C": 42
  }
}