Binary-to-text encodings, one corpus, one process. Every speed figure is divided by a baseline figure measured beside it; sizes are exact. Weight the corpus the way your traffic actually looks — the ranking is a claim about whose data is being encoded, and only you know that.
loading…
Only the ratios matter. A category at zero is still measured, just not counted. Where the same category appears in more than one corpus group it is listed once per group, so short samples can be weighted apart from large ones.
How much each factor matters to you, one slider each. They are not a see-saw: two factors both at 100 are equally important to each other, and a third at 40 is worth less than either — which is the only way to say that readable output matters to you as well as size, rather than instead of it. A factor at zero drops out of the score and out of the chart.
Size and speed come out of the corpus. These three do not, and a benchmark that quietly invented numbers for them would be worse than one that left them out. So each is a small set of tiers, every tier says what puts a codec in it, and you can disagree with a tier without having to doubt a measurement. A codec this page does not know lands in the bottom tier of each — the safe direction for something nobody has heard of.
The dashed line is the baseline codec.
The whisker is the interquartile range across rounds: how much of the bar is the machine rather than the code. Bars too close to separate are marked tied in the table below.