Hardware fingerprinting¶

The hardware fingerprint identifies the machine that produced a benchmark result. It is composed of stable identifiers from the BIOS, the GPUs, the CPU, the memory, the BIOS settings, the driver, CUDA, and NCCL. The composition is SHA-256'd into a single 64-character hex digest for fast equality checks.

Why this matters¶

Two H100 nodes can produce very different numbers depending on PCIe topology, NUMA layout, BIOS settings, and driver version. Without a fingerprint, "ran on H100" tells you almost nothing. With a fingerprint, you can prove two results came from the same configuration, or you can show that a difference is explained by a configuration change.

Composition¶

The fingerprint body is a canonical JSON object with the following fields:

fingerprint_data = {
    "dmi_uuid": dmi_uuid,
    "gpus": sorted([{
        "model": gpu.model,
        "pci_id": gpu.pci_id,
        "serial": gpu.serial,
        "vbios": gpu.vbios,
    } for gpu in gpus], key=lambda g: g["pci_id"]),
    "cpu": {"model": cpu.model, "microcode": cpu.microcode},
    "memory": {"channels": memory.channels, "speed_mts": memory.speed_mts, "ecc": memory.ecc},
    "bios": {"version": bios.version, "resizable_bar": bios.resizable_bar, "above_4g": bios.above_4g},
    "numa": numa_topology_canonical(),
    "driver": gpu_driver_version,
    "cuda": cuda_version,
    "nccl": nccl_version,
}
canonical = json.dumps(fingerprint_data, sort_keys=True, separators=(",", ":"))
fingerprint_sha256 = hashlib.sha256(canonical.encode()).hexdigest()

Fields¶

Field	Source	Why it is in the fingerprint
`dmi_uuid`	`/sys/class/dmi/id/product_uuid`	Stable per-chassis identifier.
`gpus[].model`	NVML `nvmlDeviceGetName`	Distinguishes H100 SXM5 80GB from H100 PCIe 80GB.
`gpus[].pci_id`	`lspci` / NVML	Topology matters for NVLink/PCIe paths.
`gpus[].serial`	NVML `nvmlDeviceGetSerial`	Detects board swaps.
`gpus[].vbios`	NVML `nvmlDeviceGetVbiosVersion`	VBIOS revisions change clocks.
`cpu.model`, `cpu.microcode`	`/proc/cpuinfo`, `microcode_ctl`	Microcode patches change cache/timing behavior.
`memory.channels`, `speed_mts`, `ecc`	DMI / `dmidecode`	Bandwidth-bound workloads care.
`bios.version`	DMI	Vendor BIOS updates retune memory and PCIe.
`bios.resizable_bar`, `bios.above_4g`	DMI	Both affect host-to-device transfer.
`numa`	`/sys/devices/system/node/`	NUMA layout changes effective bandwidth.
`driver`	NVML	The NVIDIA kernel driver.
`cuda`	`nvcc --version` / runtime	The CUDA toolkit runtime.
`nccl`	NCCL library	Collective communication library.

Worked example¶

bench doctor --strict

The diagnostic prints the same fields the harness will hash. After a successful bench run, you can find the fingerprint inside the envelope:

jq .hardware_fingerprint ~/.cache/inferencebench/runs/latest/envelope.json

Expected output (truncated):

{
  "fingerprint_sha256": "8b1a9c2f...",
  "dmi_uuid": "d3b07384-d9a8-4e0d-bf63-1c2e3f4a5b6c",
  "gpus": [
    {
      "model": "H100-SXM5-80GB",
      "pci_id": "0000:1b:00.0",
      "serial": "1234567890",
      "vbios": "96.00.5E.00.01"
    }
  ],
  "cpu": {"model": "AMD EPYC 9354", "microcode": "0xa101144"},
  "memory": {"channels": 12, "speed_mts": 4800, "ecc": true},
  "bios": {"version": "1.3.0", "resizable_bar": true, "above_4g": true},
  "driver": "560.35.03",
  "cuda": "12.6",
  "nccl": "2.22.3"
}

When fingerprints differ¶

Two envelopes with different fingerprint_sha256 values are not directly comparable. bench compare will surface the diff and refuse to draw a Pareto frontier across mismatched fingerprints unless you pass --allow-fingerprint-mismatch (Phase 2).