Every log behind the parfast benchmark page, the JSON the page is built from, and the harnesses that produced all of it. Each figure on that page is a median of timings written to one of these files, so any of them can be traced back to the line it came from.
Machines are named by their silicon rather than by a hostname: intel-i5-10600kf, intel-core-ultra-9-386h, amd-ryzen-9800x3d, apple-m5-max, apple-m3-ultra, apple-m1-ultra-64gb and apple-m1-ultra-128gb. The BOX line at the head of each round carries the core count, the memory and the operating system, which is what a figure has to be read against. Working directories read <rig>: which drive a round ran on says nothing about the measurement.
| Round | File |
|---|---|
| Intel i5-10600KF, 15% ladder | i5lad.log |
| Intel Core Ultra 9 386H, 15% ladder | zenlad.log |
| Apple M3 Ultra, 15% ladder | applad.log |
| Apple M5 Max, 15% ladder | m5lad.log |
| Intel i5-10600KF, full range | rounds/intel-fullrange-2026-09-11/i5full-rep1.log |
| Intel Core Ultra 9 386H, full range | rounds/intel-fullrange-2026-09-11/zfull.log |
| Apple M5 Max, full range | rounds/m5-full-range-2026-09-11.log |
| Apple M3 Ultra, full range, turbo uncapped | m3full-uncapped.log |
| Apple M3 Ultra, create sweep | rounds/msweep-m3ultra-uncapped-2026-09-11/mswp.log |
| Core Ultra 9 386H, create sweep with a memory-capped arm | rounds/zswp2-coreultra9-2026-09-12/zswp2.log |
| 25% and 30% redundancy | rounds/mred-m3ultra-2026-09-12/mred.log |
| Fast mode, per class of machine | fastmode-rounds-2026-09-12/README.md |
Banked rounds, one directory or file per round, including the ones that were aborted or came back partial. A round whose name says ABORTED or PARTIAL is kept because throwing away the legs that went wrong is how a benchmark stops being checkable, not because anything on the page quotes it.
The fast-mode cross-class rounds: eleven depths on the two x86 classes, nine half-step depths on three Apple parts, each with an A/A control.
The harnesses themselves. Every leg on this site was run by one of these, unattended, and each refuses to time a leg on a box that is not quiet.
| File | Size |
|---|---|
| harness/after.ps1 | 2.8 KB |
| harness/deadline.ps1 | 5.9 KB |
| harness/detach.py | 0.9 KB |
| harness/detachq.py | 1.0 KB |
| harness/field.ps1 | 4.5 KB |
| harness/fieldcore.ps1 | 13.3 KB |
| harness/full.ps1 | 12.8 KB |
| harness/iswp.ps1 | 0.3 KB |
| harness/jcross.ps1 | 29.3 KB |
| harness/jcross.py | 33.4 KB |
| harness/jlaunch.py | 1.1 KB |
| harness/jsmoke-avx2.ps1 | 0.1 KB |
| harness/jsmoke.ps1 | 6.4 KB |
| harness/jsum.py | 15.2 KB |
| harness/jwait.ps1 | 2.5 KB |
| harness/lad2.ps1 | 7.9 KB |
| harness/mcross.py | 15.4 KB |
| harness/mfast.py | 14.7 KB |
| harness/mqueue.py | 3.4 KB |
| harness/mred.py | 11.9 KB |
| harness/nband.py | 1.4 KB |
| harness/pdrv.py | 32.3 KB |
| harness/plib.ps1 | 32.0 KB |
| harness/repstop.ps1 | 5.2 KB |
| harness/runqueue3.ps1 | 3.9 KB |
| harness/runqueue4.ps1 | 4.3 KB |
| harness/s2sum.py | 23.4 KB |
| harness/s2x86.ps1 | 4.6 KB |
| harness/s2x86low.ps1 | 2.2 KB |
| harness/s2x86p2.ps1 | 2.1 KB |
| harness/swpcore.ps1 | 5.5 KB |
| harness/waitquiet.py | 1.4 KB |
| harness/wfast.ps1 | 11.6 KB |
| harness/wlaunch.ps1 | 7.0 KB |
| harness/wquiet.ps1 | 2.3 KB |
| harness/zen5.ps1 | 1.3 KB |
| harness/zen5aa.ps1 | 2.2 KB |
| harness/zen5rev.ps1 | 3.5 KB |
| harness/zgfni.ps1 | 25.0 KB |
The 15% ladders and the uncapped full range, read under these names by the extractor, plus the JSON the page is built from.
| File | Size |
|---|---|
| applad.log | 20.1 KB |
| bench-all.json | 65.9 KB |
| field100.json | 1.1 KB |
| field7.json | 1.0 KB |
| i5create.json | 0.2 KB |
| i5lad.log | 26.6 KB |
| m3full-uncapped.log | 19.4 KB |
| m5lad.log | 20.0 KB |
| protection.json | 2.6 KB |
| redundancy.json | 2.0 KB |
| verify.json | 1.8 KB |
| vfy-apple.json | 2.6 KB |
| zenlad.log | 25.0 KB |
98 files, 1.9 MB in all.