Tier 2 — Parametric¶
What it is¶
The default tier and the one most production traffic returns. A parametric model calibrated against ecoinvent v3.10 and ADEME Base Empreinte 2024, computed via BoaviztAPI 0.7.x with model size, hardware, datacentre PUE, live grid intensity, and embodied amortisation as inputs.
Inputs¶
| Input | Source | Update cadence |
|---|---|---|
| Model parameter count and architecture | Internal model registry | Per-release |
| Hardware specification (GPU, host CPU, DRAM) | Per-region inventory (Scaleway, atNorth, OVH) | Per-quarter |
| Datacentre PUE | Provider disclosure | Annual |
| Live grid carbon intensity | ENTSO-E + regional ISO/TSO (see grid sources) | Sub-hourly |
| Embodied carbon amortisation | EcoLogits 2025 update + ecoinvent v3.10 | Per-release |
| Batch size estimate | Trailing-quarter telemetry | Per-quarter |
Calculation¶
Roughly:
energy_per_query = (
parametric_gpu_energy(model_size, hardware, batch)
+ host_energy_share(model_size, hardware, batch)
+ idle_share(provisioned_concurrency)
) * pue
co2e_per_query = energy_per_query * grid_intensity_at_request_time
+ embodied_amortisation_per_query
water_per_query = energy_per_query * water_intensity_for_region
Each line carries a probability distribution rather than a point. The outputs are computed by Monte Carlo log-normal propagation (10,000 trials per query for tier-2; results cached and reused for repeated parameter sets within a calibration window).
Boundary¶
- Narrow: accelerator only. This is what some vendors quote in marketing.
- Comprehensive: accelerator + host CPU/DRAM + provisioned idle + datacentre PUE. This is what an assurance partner accepts and what we report by default.
The narrow ratio is typically 0.40 to 0.45 of the comprehensive number; the comprehensive boundary is the realistic operational footprint per request.
Output¶
{
"tier": "parametric",
"tier_id": "02",
"co2e": {
"median_g": 1.07,
"ci90": {"low": 0.71, "high": 1.58},
"boundary": "comprehensive"
},
"water": {"median_ml": 41, "ci90": {"low": 32, "high": 50}},
"energy": {"wh": 0.231, "narrow_wh": 0.092, "comprehensive_wh": 0.231},
"pedigree": [2, 2, 1, 1, 2]
}
Pedigree expectations¶
Tier 2 receipts typically score [2, 2, 1, 1, 2]:
- Reliability (2): verified data partly based on assumptions
- Completeness (2): representative data from > 50% of relevant sites
- Temporal (1): less than 3 years old
- Geographic (1): data from area under study
- Technological (2): data from processes and materials under study, but with somewhat different technology
The pedigree improves substantially over tier 1 in the temporal and geographic axes because we have live grid intensity at calculation time and a per-region inventory.
Calibration¶
The 90% interval reported at tier 2 is conformal-prediction calibrated against tier-3 measurements from the prior quarter. This guarantees marginal coverage on the calibration distribution: the truth has historically fallen inside the interval at least 90% of the time on the calibration workload.
The calibration is re-fit quarterly. Re-calibration events are noted in the methodology changelog.
Where this is implemented¶
methodology/tier2/parametric.py
Citations¶
- Boavizta. BoaviztAPI 0.7.x. github.com/boavizta/boaviztapi.
- ecoinvent association. ecoinvent v3.10. Zürich, 2024.
- ADEME. Base Empreinte 2024. base-empreinte.ademe.fr.
- Lloyd & Ries (2007). Characterizing, Propagating, and Analyzing Uncertainty in Life-Cycle Assessment. J. Industrial Ecology 11(1).