{
  "n_samples": 800,
  "wall_time_s": 2.7506682409439236,
  "throughput_samples_per_s": 290.8384181312504,
  "gpu_total_kernel_ms": 14.488959982991219,
  "gpu_ms_per_sample": 0.018111199978739022,
  "gpu_avg_ms_per_batch": 0.28977919965982435,
  "gpu_peak_mem_MB": 87.7998046875,
  "telemetry_pre": {
    "cpu_freq_current_MHz": 1441.831325892857,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 6.8,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 0,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1697.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 194.961,
    "gpu_temp_C": 31
  },
  "telemetry_post": {
    "cpu_freq_current_MHz": 1204.1225580357145,
    "cpu_freq_max_MHz": 4000.0,
    "cpu_util_pct": 6.3,
    "cpu_count_logical": 224,
    "cpu_count_physical": 112,
    "gpu_util_pct": 0,
    "gpu_mem_util_pct": 0,
    "gpu_mem_used_MB": 1713.0625,
    "gpu_mem_total_MB": 183359.0,
    "gpu_sm_clock_MHz": 1965,
    "gpu_mem_clock_MHz": 3996,
    "gpu_power_W": 194.966,
    "gpu_temp_C": 31
  }
}