Bbenchslop

Methodology

benchslop collects published benchmark results for CPUs, GPUs and SoCs from public databases, reduces them to robust per-chip statistics and combines them into one index per category. Nothing is entered by hand: every number traces back to a source.

Per-benchmark statistics

Indices

Within a category every result is modelled on a log scale as chip performance + benchmark scale and fitted by weighted least squares. With complete data this is exactly the geometric mean of each benchmark's ratio to the reference chip; with gaps, a benchmark's scale is pinned by the chips it shares with others. Each benchmark family counts once per chip, however many versions it has. A chip needs at least 2 benchmark families to be indexed; fewer than 3 is flagged as low confidence. GPU raster is the exception: most GPUs only have PassMark G3D, so single-benchmark GPUs are indexed (and flagged).

Index epoch 1: the reference values stay fixed within an epoch, so an index of 100 always means twice the reference chip.

Source-specific rules

Overall indices

Which chips the ranking charts show

The chart on each ranking page draws 20 chips (12 on a phone) among the ones matching the filters: the reference chip, the 3 best (2 on a phone), the 2 worst (1) among the 40 most benchmarked, each brand's most benchmarked (half the chart, split evenly between the brands), and then the most benchmarked overall. While desktop chips are shown, laptop chips get at most 4 places and phone chips 2 (2 and 1 on a small screen), since phones upload the most runs. Popularity is uploads per month over the last 12 months: Geekbench runs for CPUs and SoCs, Blender runs for GPUs. The chart follows the table's sort (index, per $100 or per watt). Chips can be added, removed or cleared with Chips or the table's tick boxes; that choice is kept in the browser and the link, and applies to every CPU (or GPU) tab.

The Pareto view plots every chip with a price against its index, price on a log axis across. Sorted per watt, it plots power instead, with the index across and power up a log axis. The frontier joins the chips no other chip beats on both, cheapest first. Prices are street prices, or launch MSRP for chips from the last 3 years; desktop parts only.

Gaming

Results read from chart images

Tom's Hardware review charts and Gamers Nexus charts (their mega-chart pages and frames from their review videos) are only published as images. Those numbers are read by a vision model and carry anextracted badge.

Prices, performance per dollar and per watt

Over time

Chip pages chart results over the last 3 years. A series is monthly when each month has at least 5results; otherwise it's quarterly under the same rule, and it isn't shown without two such periods. Benchmark series are run medians by upload month. The index series moves the current index by the average change of the chip's run-based results in that period, so it shows drift from new versions, drivers and firmware; sources that only publish averages have no history. Price series are the median of the daily price snapshots.

Sources

Only derived data is published (medians, percentiles, counts and indices); raw runs stay on the server. The full dataset is available as SQLite.