{
  "report_date": "2026-09-01",
  "hardware": {
    "gpu_name": "NVIDIA A100-SXM4-40GB",
    "gpu_memory_bytes": 42405855232,
    "compute_capability": [8, 0],
    "power_limit_W": 400.0,
    "driver_version": "580.95.05",
    "torch_version": "2.5.1+cu124",
    "cuda_version": "12.4",
    "cupy_version": "13.6.0"
  },
  "measurement": {
    "authoritative_counter": "nvmlDeviceGetTotalEnergyConsumption",
    "counter_unit": "millijoules",
    "reported_quantity": "GPU board energy above paired loaded-idle baseline",
    "scope": "GPU board including HBM; excludes host and facility",
    "trials_per_stage": 5,
    "target_seconds_per_trial": 5.0,
    "idle_window_seconds": 5.0,
    "nvidia_smi_supporting_telemetry_interval_ms": 100,
    "synchronization": "device synchronization at aggregate boundaries",
    "excluded": [
      "Modal startup and image pull",
      "CUDA and library initialization",
      "allocation and warmup",
      "host-to-device transfer",
      "datatype conversion, quantization, and bit packing",
      "CPU, host RAM, cooling, and facility overhead"
    ]
  },
  "gemm_8192": {
    "timestamp_utc": "2026-09-02T04:41:15Z",
    "source_artifact": "matmul/a100_8192_dtype_energy_results.json",
    "source_artifact_sha256": "fc5b20621d8eea98f848ed70a8a21d5bfa736a5638fd7f24d9b1c0de26c74aa6",
    "shape": [8192, 8192, 8192],
    "MAC": 549755813888,
    "conventional_operations": 1099511627776,
    "stages": {
      "strict_fp32": {
        "semantics": "FP32 x FP32 -> FP32; TF32 disabled; cuBLAS",
        "aggregate_repetitions": 435,
        "seconds_per_GEMM": 0.05794544174942529,
        "raw_board_energy_J_per_GEMM": 19.693367816091953,
        "idle_baseline_W": 67.92575465007738,
        "idle_adjusted_board_energy_J_per_GEMM": 15.75737995673014,
        "idle_adjusted_baseline_range_J": [15.649922490504519, 15.86483742295576],
        "trial_energy_stdev_J": 0.4767938572220191
      },
      "fp16_tensor_core": {
        "semantics": "FP16 x FP16 -> FP16; cuBLAS Tensor Core path",
        "aggregate_repetitions": 5805,
        "seconds_per_GEMM": 0.004362116411197246,
        "raw_board_energy_J_per_GEMM": 1.7315273040482346,
        "idle_baseline_W": 70.02403715665636,
        "idle_adjusted_board_energy_J_per_GEMM": 1.4260743023888982,
        "idle_adjusted_baseline_range_J": [1.4201229387177092, 1.4320256660600867],
        "trial_energy_stdev_J": 0.014895978187781414
      },
      "int8_tensor_core": {
        "semantics": "signed INT8 x signed INT8 -> INT32; torch._int_mm Tensor Core path",
        "aggregate_repetitions": 9055,
        "seconds_per_GEMM": 0.002778655073881832,
        "raw_board_energy_J_per_GEMM": 1.1010936499171728,
        "idle_baseline_W": 70.90474813364872,
        "idle_adjusted_board_energy_J_per_GEMM": 0.9040738117532965,
        "idle_adjusted_baseline_range_J": [0.9012033779795702, 0.9069442455270227],
        "trial_energy_stdev_J": 0.010205490085370857
      },
      "b1_and_popcount_tensor_core": {
        "semantics": "packed B1 AND + population count -> INT32; native SM80 WMMA BMMA",
        "aggregate_repetitions": 8510,
        "seconds_per_GEMM": 0.0029437698970622805,
        "raw_board_energy_J_per_GEMM": 0.8542057579318448,
        "idle_baseline_W": 70.01477899648175,
        "idle_adjusted_board_energy_J_per_GEMM": 0.6480983591725333,
        "idle_adjusted_baseline_range_J": [0.6456478051091342, 0.6505489132359326],
        "trial_energy_stdev_J": 0.02046552860227184
      }
    }
  },
  "ciresan_mnist_int8_inference": {
    "timestamp_utc": "2026-09-02T05:15:16Z",
    "modal_app_id": "ap-6tbqDsQQ2rxZ6Z32HEcN05",
    "source_artifact": "ciresan_a100/a100_40gb_int8_batch16_results.json",
    "source_artifact_sha256": "73b221fb089fb1216ab8bb160eda90e3e3dca60a89f99f5b4c6f443a558cf037",
    "logical_dimensions": [784, 2500, 2000, 1500, 1000, 500, 10],
    "physical_width_padded_dimensions": [784, 2512, 2000, 1504, 1008, 512, 16],
    "examples": 60000,
    "logical_MAC_per_example": 11965000,
    "logical_MAC_per_epoch": 717900000000,
    "weights": "deterministic synthetic dense signed INT8",
    "input": "real MNIST training images, uint8 minus 128 then signed INT8",
    "epilogue": "INT32 bias + ReLU + saturation to [0,127] + INT8 write after every layer",
    "stages": {
      "logical_batch16": {
        "calls_per_epoch": 3750,
        "physical_rows_per_call": 32,
        "row_padding_reason": "torch._int_mm rejects M=16 on this A100/PyTorch build",
        "physical_padded_MAC_per_epoch": 1445007360000,
        "aggregate_repetitions": 35,
        "seconds_per_epoch": 0.780364251971429,
        "raw_board_energy_J_per_epoch": 61.033371428571435,
        "idle_baseline_W": 62.69807690575121,
        "idle_adjusted_board_energy_J_per_epoch": 12.106033543967767,
        "idle_adjusted_baseline_range_J": [11.42979997406332, 12.782267113872201],
        "trial_energy_stdev_J": 0.5154639408925056
      },
      "single_full_batch60000": {
        "calls_per_epoch": 1,
        "physical_padded_MAC_per_epoch": 722503680000,
        "aggregate_repetitions": 3260,
        "seconds_per_epoch": 0.007773678443865031,
        "raw_board_energy_J_per_epoch": 3.0642294478527603,
        "idle_baseline_W": 63.641640426835274,
        "idle_adjusted_board_energy_J_per_epoch": 2.5694997995344617,
        "idle_adjusted_baseline_range_J": [2.5486108185709284, 2.590388780497995],
        "trial_energy_stdev_J": 0.041068117246861116
      }
    },
    "excluded_additionally": [
      "labels, loss, accuracy, backward, and optimizer",
      "initial uint8-to-signed-int8 conversion",
      "model and workspace allocation"
    ]
  }
}