{
  "n_samples": 800,
  "wall_time_s": 2.1655724809970707,
  "throughput_samples_per_s": 369.41732822152636,
  "gpu_total_kernel_ms": 20.288064047694206,
  "gpu_ms_per_sample": 0.02536008005961776,
  "gpu_avg_ms_per_batch": 0.40576128095388414,
  "gpu_peak_mem_MB": 55.22314453125,
  "telemetry_pre": {
    "cpu_freq_current_MHz": 1046.6589866071429,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 21.2,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 0,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1679.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 193.475,
    "gpu_temp_C": 31
  },
  "telemetry_post": {
    "cpu_freq_current_MHz": 1074.5570491071428,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 1.8,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 0,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1679.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 193.35,
    "gpu_temp_C": 31
  }
}