Skip to content

Engine Coefficients

GRADE aggregates engine scores using published coefficients. The coefficients reflect the relative contribution of each dimension to overall AI compute infrastructure efficiency. Coefficients are versioned — changes are documented with rationale.

EngineWeightRationale
ACE0.35GPU utilization is the primary measure of compute efficiency
PACE0.25Scheduler policy determines whether utilization is achievable
COOL0.20Cooling overhead is the largest non-compute energy consumer
CORE0.12Hardware fit affects efficiency ceiling; embodied carbon is a lifecycle factor
FLUX0.08Carbon accounting methodology matters; weight reflects current data availability

Coefficients sum to 1.00. When an engine is excluded from an assessment, remaining engine scores are normalized proportionally before aggregation.

If COOL and FLUX are excluded (a common pattern for first assessments), the remaining engines — ACE, PACE, CORE — are re-weighted proportionally:

  • ACE: 0.35 / 0.72 = 0.486
  • PACE: 0.25 / 0.72 = 0.347
  • CORE: 0.12 / 0.72 = 0.167

The composite is computed against the normalized weights. The certification report discloses which engines were included and the normalized weights applied.

GPU utilization directly measures whether allocated compute is doing useful work. A cluster where 89% of jobs run below the 40% utilization threshold is wasting the majority of its purchased GPU capacity, regardless of how well-cooled or carbon-accounted it is. ACE is weighted highest because it is the most direct measure of the fundamental problem PTL is designed to address.

Carbon accounting methodology is important. It is also the dimension where organizational control is most constrained by geography, utility availability, and regulatory environment. An organization in a region with no viable PPA market cannot easily move from 0.50 to 1.00 on FLUX in one assessment cycle. The coefficient reflects this constraint — FLUX is measured and reported, but not allowed to dominate the composite.

Current coefficients: v2.0, published 2026-01-15. Changes from v1.0: CORE weight increased from 0.08 to 0.12 (Blackwell embodied carbon significance); FLUX weight decreased from 0.12 to 0.08 (geographic constraint acknowledgment). Previous version coefficients are archived in the ptl-methodology repository.