Benchmarks
25 published runs. The newest is
14 days ago — benchmarks do not run on this site's
builder (no GPU, and the quick profile alone is seven minutes), so these are as fresh as the last
npm run benchmark:publish.
Intel Core i9-9900K 8c/16t · 62.6 GiB · CoffeeLake-S GT2 [UHD Graphics 630] + Ellesmere [Radeon RX 470/480/570/570X/580/580X/590] (8.0 GiB)
Machine id 5cf3f79c. Only runs sharing this id are comparable.
A run marked new harness was measured by a benchmark harness the previous run of its profile did not use, so reading the two as a before and an after is reading two different instruments. The comparison pages show those rows as breaks rather than as changes.
Across commits: all (7 runs) · algorithms-gpu (5 runs) · renderer (7 runs) · quick (6 runs)
| run | profile | commit | geo-mean vs v3 | sampling | duration | note |
|---|---|---|---|---|---|---|
| 2026-09-02 02:03 | all | 9ed49abc |
10.70× | median of 3 | 45.4 min | round 113 verification run at 9ed49abc, after the 113.2 predicate-filter fix: 269 pairs, geo-mean 10.7x, minimum 1.02x, zero v4-slower; against 13 Aug zero screened regressions and the three filter(fn) rows as the improvements |
| 2026-09-02 01:47 | algorithms-gpu | 9ed49abc |
– | 1 run | 16.3 min | round 113 verification: executor sweep on the RX 580 at 9ed49abc — no src/algorithms change since 13 Aug; the GPU side reads 20-50% faster at n=512 (driver/box), the cpu control flat |
| 2026-09-02 01:25 | renderer new harness | 9ed49abc |
31.21× | 1 run | 22.3 min | round 113 verification: renderer scenes on the RX 580 at 9ed49abc — the first published renderer run since the 87.2/95.1 harness change, so every row against 13 Aug is an epoch break by design |
| 2026-09-01 23:46 | all new harness | 5107c36e |
10.61× | median of 3 | 45.4 min | round 113 review run at 5107c36e (before the 113.2 filter fix): the full core sweep, medians of 3, Node 24 — the first published run after the fingerprint re-stamp, so it is the first to compare against 13 Aug |
| 2026-08-13 01:30 | algorithms-gpu new harness | 0d3561de |
– | 1 run | 16.4 min | post-merge baseline at 0d3561de: executor sweep incl. the round 69/70 families |
| 2026-08-13 01:09 | renderer | 0d3561de |
32.52× | 1 run | 21.1 min | post-merge baseline at 0d3561de: renderer scenes on the RX 580 |
| 2026-08-13 00:24 | all new harness | 0d3561de |
10.70× | median of 3 | 44.9 min | post-merge baseline at 0d3561de: full core sweep, medians of 3 |
| 2026-08-12 23:08 | algorithms-gpu new harness | c2d24c2e |
– | 1 run | 4.9 min | round 68 status refresh: cpu vs gpu executor sweep on the RX 580 |
| 2026-08-12 22:46 | renderer new harness | c2d24c2e |
32.29× | 1 run | 21.1 min | round 68 status refresh: renderer scenes on the RX 580 |
| 2026-08-12 15:45 | all new harness | c2d24c2e |
10.70× | median of 3 | 45.0 min | round 68 status refresh: full core sweep, medians of 3 |
| 2026-08-11 15:16 | quick | 54062c08 |
19.34× | median of 3 | 22.6 min | round 65.12: replicate B, same commit and harness as A. Any change between the two is noise by construction — the pair exists so the comparison has a measured band to screen against, and so the screen's own false-positive rate is on the record (1 of 217 rows here, against 6 under a flat ±10%). |
| 2026-08-11 14:54 | quick new harness | 54062c08 |
19.37× | median of 3 | 22.6 min | round 65.12: first --repeat 3 run — each row is the median of three processes and carries the band they spanned. Replicate A of a same-commit pair published to give the new harness epoch its measured noise floor. |
| 2026-08-10 22:13 | algorithms-gpu new harness | 44e55dc9 |
– | 1 run | 4.9 min | round 65.10: executor sweep with the pageRankDense family; pageRank and hierarchical now route 'auto' to the CPU (sparse iteration / flat merges), the GPU keeping the seven families it wins |
| 2026-08-10 21:55 | all | 44e55dc9 |
10.80× | 1 run | 18.4 min | round 65.10: all profile after sparse pageRank and the flat hierarchical merge engine — both CPU-executor changes, so the v3-comparative rows move |
| 2026-08-10 19:05 | algorithms-gpu | 1f324088 |
– | 1 run | 5.0 min | round 65.9: first published run of the algorithms-gpu profile — executor cpu vs gpu per family per size on the amd gcn-4 adapter, after the 65.8 kernel perf pass |
| 2026-08-10 18:46 | all new harness | 1f324088 |
10.42× | 1 run | 18.4 min | round 65 baseline: all profile after 65.1-65.9b — style-bundle measures again for the first time since round 42 |
| 2026-08-10 18:18 | quick new harness | d4ff0730 |
18.89× | 1 run | 7.5 min | round 65 baseline: quick profile after the async algorithm tier (65.1) and the kernel perf pass (65.8) — the algorithm rows here are the CPU executor, which 65.8 also sped up (AP build fix, typed median) |
| 2026-08-10 10:18 | renderer | c182dd0c |
27.87× | 1 run | 20.1 min | |
| 2026-08-10 04:38 | all | cf5727b4 |
10.92× | 1 run | 17.9 min | |
| 2026-08-09 23:08 | quick | 6df994f1 |
12.60× | 1 run | 7.1 min | |
| 2026-08-09 16:52 | renderer new harness | e37d2444 dirty |
27.37× | 1 run | 19.9 min | round 60 baseline: all 12 renderer scenes after the arrow trim (56), state conditions (57.1d) and the force rebuild (59). The dirty flag is the archive itself: the quick run published minutes earlier sat uncommitted in benchmark/published/ while this measured; no source differed from e37d2444 |
| 2026-08-09 16:39 | quick new harness | e37d2444 |
11.32× | 1 run | 7.1 min | round 60 baseline: quick profile after the round-59 force rebuild and round-60 tooling |
| 2026-08-06 23:41 | renderer new harness | 3c9c3e33 |
24.69× | 1 run | 11.3 min | round 55 baseline: the arrow scenes and their arrows-off control, before the gap/trim port |
| 2026-08-05 18:38 | renderer | 365d4d5a |
27.45× | 1 run | 14.0 min | round 46.5 baseline: all 8 renderer scenes, incl. the ndex scene recovered this round |
| 2026-08-05 18:11 | quick | dc6d3505 |
12.99× | 1 run | 7.1 min | round 46.5 baseline: CPU/collection quick profile |