CORE — Capital and Operations Resource Evaluator

The Capital and Operations Resource Evaluator (CORE) grades the hardware layer of your AI infrastructure — whether the GPUs you operate are the right hardware for the workloads you run, how old the fleet is relative to the GPU refresh cycle, and how efficiently embodied carbon is amortized across the compute capacity those GPUs deliver. CORE identifies whether hardware decisions are creating a structural ceiling on what ACE and PACE can achieve.

Version

engine v0.1.0fit matrix update 2026-05-09MITchecks passing

Primary metric

core_efficiency_score — a number from 0.0 to 1.0 composed of three fleet-weighted components:

• Hardware fit (40%) — how well each GPU model matches the declared primary workload type
• Fleet age (35%) — GPU age in years, anchored to published refresh-cycle references
• Embodied carbon (25%) — carbon efficiency normalized by compute capacity delivered (kg CO₂e per TFLOP/s)

Formula

core_efficiency_score =
  hardware_fit_score    × 0.40 +
  fleet_age_score       × 0.35 +
  embodied_carbon_score × 0.25

All three component scores are GPU-count-weighted fleet averages. A fleet with 800 H100s and 200 A100s weights H100 at 80% and A100 at 20%.

Hardware-workload fit

Scores how well each GPU model suits the declared primary workload. Derived from MLPerf Training v3.1 and MLPerf Inference v4.0 benchmark results (mlcommons.org/results), normalized so the best available result per workload = 1.00. AMD and Intel scores from their respective MLPerf submissions.

Selected fit scores:

GPU	training_large	training_small	inference	mixed
B200	1.00	0.92	0.95	0.97
GB200	1.00	0.90	0.95	0.97
B100	0.95	0.90	0.92	0.93
H200	0.90	0.89	0.93	0.91
H100	0.88	0.88	0.90	0.89
MI300X	0.85	0.88	0.90	0.87
Gaudi3	0.72	0.75	0.78	0.75
A100	0.52	0.72	0.60	0.62
L40S	0.42	0.72	0.83	0.68
Gaudi2	0.55	0.62	0.60	0.58
A40	0.38	0.70	0.75	0.60
RTX Pro 6000	0.40	0.72	0.88	0.65
RTX 4090	0.28	0.58	0.72	0.50
V100	0.22	0.42	0.38	0.32
unknown	0.40	0.40	0.40	0.40

Consumer GPUs (RTX 4090, RTX 3090) have no MLPerf Datacenter submission — their scores are extrapolated from published FLOP/s ratios and carry higher uncertainty.

MIG utilization multiplier

When MIG partitioning is enabled on a GPU record, the fit score is multiplied by active_instances / total_instances. Idle MIG partitions represent configured but unused compute capacity.

mig_multiplier = mig_instances_active / mig_instance_count
fit_score_effective = fit_score × mig_multiplier

Fleet age scoring

Age is computed from assessment_year − purchase_year per GPU record, then weighted by GPU count across the fleet. Scoring is anchored to published GPU hardware refresh-cycle references.

Age (years)	Score	Rationale
≤ 2	1.00	Current generation, full performance, full software support
≤ 4	0.80	One generation behind, still widely supported
≤ 6	0.60	Two generations behind, approaching EOL driver support
≤ 8	0.35	EOL or near-EOL, limited workload support
> 8	0.15	Legacy, significant efficiency penalty

Sources: public hyperscaler sustainability reports; GPU generation cadence references; DOE ESIF hardware lifecycle documentation.

Embodied carbon scoring

Embodied carbon is normalized by compute capacity delivered — kg CO₂e per TFLOP/s — rather than absolute carbon per GPU. This correctly reflects that a high-performance GPU with higher embodied carbon may be more carbon-efficient per unit of useful compute than a legacy GPU with lower absolute carbon.

kg_per_tflops = embodied_carbon_kg / tflops_per_gpu
score = 1 − (kg_per_tflops / 2.24)   [clamped 0.0 – 1.0]

Reference baseline: A100 at 700 kg CO₂e / 312 TFLOP/s = 2.24 kg/TFLOP/s → score 0.0. GPUs with better compute-per-carbon score above 0.0. GPUs below A100 density also score 0.0 (clamped).

Selected embodied carbon values and scores:

GPU	TFLOP/s (FP16)	CO₂e (kg)	kg/TFLOP/s	Score
B200	4,500	1,200	0.267	0.881
H100	1,979	700	0.354	0.842
MI300X	2,614	950	0.364	0.838
H200	1,979	950	0.480	0.786
Gaudi3	1,835	900	0.490	0.781
L40S	733	420	0.573	0.744
A100	312	700	2.244	0.0 (reference)
V100	125	500	4.000	0.0 (clamped)

Sources: NVIDIA Product Carbon Footprint reports; Dell/HPE lifecycle data; Gupta et al. 2021; Patterson et al. 2022. Blackwell values extrapolated from die-size scaling.

Worked example

Fleet: 512 H100s purchased 2023, 256 A100s purchased 2021. Primary use: training_large. Assessment year: 2026.

H100 record (512 GPUs, age = 3 years):
  fit_score       = 0.88  (H100 × training_large)
  age_score       = 0.80  (age 3 ≤ 4)
  embodied_score  = 0.842 (700 kg / 1979 TFLOP/s = 0.354 kg/TFLOP/s)

A100 record (256 GPUs, age = 5 years):
  fit_score       = 0.52  (A100 × training_large)
  age_score       = 0.60  (age 5 ≤ 6)
  embodied_score  = 0.00  (reference baseline — clamped)

Fleet-weighted averages (512 H100 + 256 A100 = 768 total):
  hardware_fit_score    = (512×0.88 + 256×0.52) / 768 = (450.6 + 133.1) / 768 = 0.760
  fleet_age_score       = (512×0.80 + 256×0.60) / 768 = (409.6 + 153.6) / 768 = 0.733
  embodied_carbon_score = (512×0.842 + 256×0.00) / 768 = 431.1 / 768 = 0.562

core_efficiency_score =
  0.760 × 0.40 +
  0.733 × 0.35 +
  0.562 × 0.25
= 0.304 + 0.257 + 0.141
= 0.702  → Grade B (Good)

The A100s drag down all three components. Refreshing the 256 A100s to H100s would push composite to ~0.85 (Grade A).

Letter grade

Grade	Condition	Label
A	score ≥ 0.80	Excellent
B	score ≥ 0.65	Good
C	score ≥ 0.50	Adequate
D	score ≥ 0.35	Poor
F	score < 0.35	Failing

Findings exported to GRADE

Metric	Unit	Description
core_efficiency_score	score_0_to_1	Weighted composite — GRADE input
hardware_fit_score	score_0_to_1	Fleet-weighted fit for declared primary workload
fleet_age_score	score_0_to_1	Fleet-weighted age band score
embodied_carbon_score	score_0_to_1	Fleet-weighted kg CO₂e per TFLOP/s score
fleet_age_years_weighted	years	GPU-count-weighted mean fleet age
hardware_efficiency_grade	grade_points_0_to_4	A=4, B=3, C=2, D=1, F=0

ACE cross-reference

CORE reads gpu_efficiency_score from an ACE output when ace_report_path is provided. This is informational — it does not affect the CORE composite formula in v0.1, but it appears in the GRADE report alongside CORE findings to give context on whether a hardware fit gap is being exposed by actual workloads.

Input schema

CORE takes a JSON file with a hardware fleet list and a workload profile. The fleet is a list of HardwareRecord objects — one per distinct GPU model and purchase cohort.

Field	Type	Required	Description
organization	string	yes	Organization name
period	string	yes	Assessment year (YYYY)
scheduler	string	yes	slurm, kubernetes, or unknown
hardware_fleet	list	yes	List of hardware records (see below)
workload_profile	object	yes	Workload characteristics
ace_report_path	string	no	Path to ACE ptl_output for cross-reference

Hardware record fields:

Field	Type	Required	Description
gpu_model	string	yes	h100, a100, b200, mi300x, v100, etc.
gpu_count	int	yes	Number of GPUs in this record
purchase_year	int	yes	Year purchased or last substantially upgraded
tdp_watts	int	yes	GPU TDP in watts
embodied_carbon_kg	float	no	Total embodied CO₂e for record; uses default if absent
mig_enabled	bool	no	Whether MIG partitioning is active
mig_instance_count	int	no	Total MIG instances configured
mig_instances_active	int	no	MIG instances currently in use

Workload profile fields:

Field	Type	Required	Description
primary_use	string	yes	training_large, training_small, inference, mixed, or unknown
avg_gpu_utilization	float	yes	Average GPU utilization (0.0–1.0)
avg_job_duration_minutes	float	yes	Average job duration in minutes

CLI usage

# Score a single organization from a core_input JSON file
core analyze \
  --input core_input.json \
  --output core_output.json

# Run all synthetic hardware fleets and print grades
core demo

# Validate a ptl_output_v1.json export against the PTL schema
core validate --input core_output.json

Example core_input.json:

{
  "organization": "National AI Research Center",
  "period": "2026",
  "scheduler": "slurm",
  "hardware_fleet": [
    {
      "gpu_model": "h100",
      "gpu_count": 512,
      "purchase_year": 2023,
      "tdp_watts": 700
    },
    {
      "gpu_model": "a100",
      "gpu_count": 256,
      "purchase_year": 2021,
      "tdp_watts": 400
    }
  ],
  "workload_profile": {
    "primary_use": "training_large",
    "avg_gpu_utilization": 0.72,
    "avg_job_duration_minutes": 240
  }
}

Source

• View CORE source on GitHub
• CORE methodology document
• Report an issue