Methodology — Quanteta GPU Performance Index

What these indices are (and are not)

Every number on this site is a Quanteta index calculated by formula from a GPU's published specifications (FP16/FP32 throughput, memory bandwidth, clocks, TDP, core counts). They are not measured benchmarks: no game was run, no FPS or render time was recorded. Like Walk Score or an Elo rating, each index reduces published data to a clean, comparable 0-100 scale via a documented, reproducible formula. We publish every formula below so the numbers can be audited and recomputed.

Indices

Five calculated 0-100 indices, normalized across the whole dataset. Each is a transparent formula over published specifications (compute throughput, bandwidth, clocks, TDP) — analytical indices, not measured game/render benchmarks, and never an editorial opinion.

Overall Performance Index (QOV)

Composite 0-100 index calculated from published specifications — not a measured benchmark. Weighted blend of Compute Strength (30%), Graphics Index (30%), Power Efficiency Index (20%) and Memory Bandwidth Density (20%), normalized across the whole dataset.

Compute (AI) Strength (QAI)

0-100 index for AI and compute workloads, calculated from specs (not a measured benchmark). Derived from FP16 throughput (45%), memory bandwidth (30%) and tensor-core count (25%). Cards with under 8 GB VRAM score 0 on this axis.

Graphics Index (QGP)

0-100 index for graphics / gaming workloads, calculated from specs (not a measured benchmark). Derived from texture rate (35%), pixel rate (25%), memory bandwidth (25%) and boost clock (15%).

Power Efficiency Index (QWP)

0-100 index measuring FP32 throughput per watt of TDP, calculated from specs (not a measured benchmark). Higher means more compute per watt.

Memory Bandwidth Density (QME)

0-100 index measuring memory bandwidth relative to VRAM capacity and bus width — i.e. how dense the memory subsystem is, not overall capability. Calculated from specs (not a measured benchmark). By construction this favours cards with narrow, fast memory and small VRAM, so some older / console GPUs rank highly here. Use it for memory-subsystem analysis, not as a performance ranking.

Exact formulas

• Compute (AI) Strength (QAI) = 0.45·FP16(TFLOPS) + 0.30·(bandwidth/100) + 0.25·(min(tensor_cores/512, 1)·25); cards under 8 GB VRAM = 0. Then min-max normalized to 0-100.
• Graphics Index (QGP) = 0.35·texture_rate + 0.25·pixel_rate + 0.25·(bandwidth/10) + 0.15·(boost_clock/100). Then min-max normalized to 0-100.
• Power Efficiency Index (QWP) = FP32(TFLOPS) / TDP(W) × 1000. Then min-max normalized to 0-100.
• Memory Bandwidth Density (QME) = 0.5·(bandwidth/VRAM_GB) + 0.5·(bandwidth/(bus_bits/8)). Then min-max normalized to 0-100.
• Overall Performance Index (QOV) = 0.30·QAI + 0.30·QGP + 0.20·QWP + 0.20·QME (over available components).

A GPU is indexed on a given axis only when all inputs for that axis are present; otherwise the axis is omitted rather than imputed. The short codes (QOV, QAI, QGP, QWP, QME) are compact column labels for the indices named above.

Percentile ("top X%")

Each index on a GPU page is shown with a percentile such as "top 1%". The population for a given axis is every GPU in this database that has an index on that same axis — not a time-windowed or vendor-filtered subset. As of the current build that is 2455 GPUs for the Overall Performance Index (QOV) axis, 721 for Compute/AI Strength (QAI) and 2435 for the Graphics Index (QGP). "Top X%" means the GPU's index is at or above the (100−X)th percentile of that population. Counts shift as the dataset grows.

Data Sources

GPU Data

RightNow GPU Database — public GPU specifications dataset.

Update Frequency

Collected periodically via automated collectors. Pages are regenerated only when underlying data changes. The "Last updated" timestamp reflects actual data changes, never cosmetic edits.

Data Quality

• Validation: incoming data is validated against expected ranges and types.
• Anomaly detection: implausible FP32 / TDP values are flagged.
• Automated collection: data is collected by scripts to eliminate manual entry errors.
• Single source of truth: JSON-LD structured data and HTML are generated from the same variables; a build-time gate fails the build on any mismatch.

Limitations & Caveats

• Data depends on the accuracy of the upstream source.
• Indices are descriptive metrics calculated from specifications — not measured benchmarks, predictions or recommendations.
• Coverage varies; not every GPU has complete data for all fields or all index axes.
• The source dataset contains no pricing, so no price or value-per-dollar metric is published.
• Memory Bandwidth Density (QME) by construction favours narrow, fast, small-VRAM memory subsystems, so some older / console GPUs rank highly on that axis. This is an honest property of the formula, not a ranking of overall capability.