What the export contains

463 values · 181 unknown · 64 not applicable · 0 not yet checked

All 81 systems have at least one numeric measurement. Documented unknowns remain beside the values they qualify.

Exports

catalog-v1.json The full export: catalog and research systems, components, measurements, contextual claims, sources, derived claims and conflicts.
catalog-summary-v1.json A compact index of names and facets for selectors and filters. It contains no figures, so a filter does not need to download the catalog.
measurements-v1.csv One row per figure, with its unit, its normalized value in the base unit, its comparability group, its confidence status and its sources.

The artifacts carry no timestamp and no build identifier. Two builds from identical inputs produce byte-identical files. A reviewer can therefore inspect a data change as a diff.

What is in it

Systems: 81
Contextual claims: 468
Components: 250
Figures: 708
Sources: 196
Derived claims: 57
Conflicts: 8

Schema and registries

The export declares the schema it conforms to and the version of every registry it was built against. Changing a definition changes its registry version and forces derived results to be recomputed and re-checked.

Schema: catalog-v1
Units: units-v1 Every unit belongs to one quantity. Machine words and bytes, and Dhrystone MIPS and native MIPS, are separate quantities, not flags on a shared one, because they must never convert into one another.
Metrics: metrics-v1 Each metric fixes what is measured, which device scopes it applies to, and whether a ratio or a logarithmic axis is meaningful for it at all.
Methods: methods-v1 A method says how a figure was obtained, and it is part of the comparability group. Who published it and whether the hardware had shipped are recorded separately, because neither changes what the number measures.
Comparability: comparability-v1 A comparability group is metric, scope, method and benchmark identity. The unit is not part of that group.

Licensing

The code is MIT. The catalog, the editorial prose and the exports above are CC BY 4.0. You may share and adapt them with credit to RetroFlops. Individual facts are not owned by anybody; the license covers the compilation: the machines selected, the figures judged acceptable, their normalization and grouping, and the prose that explains them.

The documents cited belong to their publishers, and nothing here grants any right over them. Each source record states the license its publisher applies and any usage constraint that travels with a quotation from it. SPEC results, for instance, carry the SPEC fair-use rules wherever they appear. One source states no license at all, and its record does not assume any permission.

Short extracts from vendor documentation are quoted only for verification, so readers can check a figure against its source.

Reporting an error

Report a correction with the record and the document that contradicts it. The catalog records a conflict instead of an average. If two acceptable sources disagree, both values are recorded with their evidence and an editorial decision. A documented disagreement can become part of the catalog.

Figures whose sourcing is not yet strong enough are marked provisional and take no part in any automatic comparison. A source that settles one lets the editors review it. The methodology sets out what counts as an acceptable source.