Datasets / Local AI GPU Value Index / Methodology

GPU Value Index — Methodology

Version v2026-09-04 · License CC BY 4.0

What the dataset is

One row per discrete GPU we track (Consumer, Professional, Data Center categories). Two tiers of columns live side by side:

• Spec tier — vendor-published VRAM and memory bandwidth divided by launch MSRP. A coverage proxy computable for every sourced-price GPU.
• Measured tier — independently measured llama.cpp generation throughput (tokens/s) per $1000 of launch MSRP, available only where a non-estimate benchmark row exists in a defined cohort.

Cohorts (never blended)

The two cohorts differ in model AND quantization; their numbers are never averaged, normalized against each other, or merged into a single score. Where both exist for one GPU they appear in separate columns.

Formulas

• bw_per_dollar_1000 = memory_bandwidth_gbps / msrp_usd × 1000 — GB/s per $1000.
• vram_per_dollar_1000 = vram_gb / msrp_usd × 1000 — GB per $1000.
• tok_s_per_1000__<cohort> = measured_tok_s__<cohort> / msrp_usd × 1000.

All divisions round half-up to 2 decimals. Every figure on the landing page is recomputable from the CSV download.

Evidence legend

• vendor_spec — manufacturer specification page (VRAM GB, memory bandwidth GB/s). Linked per row via the product's source URL.
• manufacturer_pricing — launch MSRP in USD at introduction. This dataset uses only launch MSRP (price_basis=launch_msrp on every row). It is not a current-street or used-market price index.
• independent_benchmark — third-party measured llama.cpp single-stream generation throughput. Only rows flagged non-estimate in our database qualify; each carries *_bench_source_url and *_bench_tested_at.

Exclusions & NULL policy

• Apple Silicon excluded. Unified memory is not VRAM; mixing it would corrupt per-GB metrics.
• No imputation. Missing MSRP ⇒ NULL dollar metrics + dollar_metric_null_reason=no_msrp_recorded. Missing benchmark ⇒ NULL measured columns. We never estimate, interpolate, or scale numbers to fill gaps.
• Estimates never enter ranked data. Rows flagged as estimates in our benchmarks database are excluded from this dataset entirely.
• Generation tokens/s only. Prompt-processing and batch variants measure different things and are out of scope.
• Data Center GPUs without recorded MSRPs appear with NULL dollar metrics for completeness.

Known limitations

The measured tier currently covers a modest number of GPUs (see coverage on the landing page). Spec-tier metrics are a proxy: token generation is largely bandwidth-bound, but real performance depends on cooling, drivers, and software stack. Cross-cohort comparisons are invalid by construction.

Refresh cadence

The exporter runs nightly and regenerates when any source table changes (products, specifications, benchmarks) or at least weekly. Each regeneration produces a new date-based version (vYYYY-MM-DD). Versioned URLs are stable citations; /latest.csv always points at the current version.

How to cite

Suggested citation:

CompareAIHardware, "Local AI GPU Value Index", dataset version v2026-09-04. Licensed under CC BY 4.0. https://compareaihardware.com/datasets/gpu-value-index

Attribution requirement: keep the linked name "CompareAIHardware" (or link the dataset URL) visible wherever the data is republished.

Column reference

See the columns array in the JSON envelope for the authoritative ordered list; each name is snake_case and stable across versions.