{
  "date": "2026-08-23",
  "device": "NVIDIA L20",
  "cuda_runtime": "13.0",
  "torch": "2.12.1+cu130",
  "model": "tiny-llama-1.1b",
  "parameters": 1100048384,
  "sequence_length": 2048,
  "precision": "bf16-mixed",
  "torch_compile": true,
  "runs": [
    {
      "micro_batch_size": 4,
      "measured_steps": 10,
      "tokens_per_second": 12274.280878357904,
      "peak_allocated_gib": 28.84268045425415,
      "peak_reserved_gib": 31.001953125,
      "status": "pass"
    },
    {
      "micro_batch_size": 6,
      "measured_steps": 60,
      "tokens_per_second": 12586.170876605223,
      "peak_allocated_gib": 36.131224155426025,
      "peak_reserved_gib": 36.294921875,
      "status": "pass"
    },
    {
      "micro_batch_size": 8,
      "status": "oom_on_second_warmup_backward",
      "memory_in_use_gib": 44.17,
      "free_mib": 215.31,
      "failed_allocation_mib": 1000
    }
  ],
  "telemetry": {
    "samples": 12,
    "sm_utilization_percent_min": 100,
    "sm_utilization_percent_max": 100,
    "power_watts_min": 343,
    "power_watts_max": 358,
    "framebuffer_mib": 37531
  },
  "selected_micro_batch_size": 6,
  "selected_gradient_accumulation": 85,
  "selected_effective_global_batch_sequences": 510,
  "estimated_compute_only_days_for_20b": 18.39
}