Per-card history
One card over time
The leaderboard keeps a single row per card, workload and release: the best run wins, and every later run is dropped. That is the right view for comparing hardware, and the wrong one for watching a card age. This page plots every run of one card instead, so a part that is slowing down shows up as a slope rather than disappearing behind its own best result.
Reading these charts
Cards are identified by the GPU UUID the driver reports, published as-is. A UUID names one physical card, not the machine it sat in, and it follows the card across releases and owners, which is what a history needs. Host details (hostnames, network information) are removed before a report is published. Cards whose report carried no usable UUID (early reports, and AMD reports that give 0x0) are grouped by model, serial number when there is one, memory size and driver instead and labelled as having no GPU ID. Two identical cards with the same driver cannot be told apart that way and share one series.
When a workload's metric changed between releases, each unit is drawn as its own series. Values in different units are not comparable, and joining them into one line would show a step that is a change of units rather than a change in the hardware.
Runs of different lengths are drawn as separate series as well. A longer run heat-soaks the card and usually scores lower than a short one on the same silicon, so mixing a 30 s, a 300 s and a 3600 s result in one line would show a decline that is not there.
A workload that Pantheon reworked on purpose is split in the same way. Eleven workloads score differently from release 1.2.3 on (the list is under Scores that changed between releases), so their results from 1.2.3 on are drawn as a separate series, labelled "from 1.2.3", apart from the earlier ones. Joined into one line, a card tested under both releases would seem to jump by tens of per cent when only the test changed.
Runs come from the same reports as the leaderboard. A single benchmark writes both a per-test record and a summary repeating it, so those are collapsed to one point; two separate runs are only merged when they share a timestamp as well as a score, at which point they are the same measurement counted twice.