a88cee9241
bench/: repro.sh (Linux, macOS) and repro.ps1 (Windows) print the machine, check the genesis pack's lottery hash vectors on the CPU and every GPU, hash for 120 s per card at 1 GiB, probe random reads at 4, 64, 256 and 1024 MiB, sweep the same sizes, prove and verify one fixture shard when a 12 GB NVIDIA card and a prover host exist, and write one JSON result (the bench table's row shape, every command, every binary's sha256, the tolerances) plus a table. make-package.sh builds igneum-repro-<tag>.tar.gz and .zip from the existing cross-build scripts; ingest.mjs checks a result against bench/reference.json and adds "reproduced externally" rows to site/miner-bench.json only when every number agrees; publish-public.sh --repro publishes next to the downloads (not deployed). The workers gained --list, --bench --pack --seconds --dataset-mib and --memprobe (CUDA from the readwidth branch, OpenCL memprobe from opencl-rdna4, Metal new) with RESULT and DEVICE lines; igneum-pow gained check-pack; the pack reader accepts a string-seed pack. Runs (docs/benchmarks/repro.md, the bench-log entry, evidence rows 4, 5, 15): the Mac under the measure lock (Metal 27.674 MH/s, Apple OpenCL 26.875, 96/96 on CPU, Metal and OpenCL, 2^24 fingerprint 25f96e7dce90bd4e, 3.40 G random reads/s, CPU verify 0.611 ms per warp); PC 2 (96/96 through NVRTC, NVIDIA OpenCL and AMD OpenCL with the same fingerprint; gfx1036 3.312 MH/s; the 5090 rows beside the app's live miner, 62.4 MH/s, not the card's; the live host proved and verified the shard through step 6, compressed 33.9 s). PC 1 went down before its slot; owed. Classes closed: the PC job parses the package script before touching a card and bench/ is in the Windows 5.1 parse job; the card-off is confirmed by nvidia-smi's compute-apps list and the prover's state read from settings.json, never api/state; PowerShell case-only variable pairs fail CI (bench/jobs/ps-case-check.sh); the package builder keeps the Mac workers outside the stage and checks every binary with file -b. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
64 lines
2.7 KiB
JSON
64 lines
2.7 KiB
JSON
{
|
|
"_about": "The team's reference numbers per card model for bench/ingest.mjs: the bench-log entry each comes from (docs/bench-log.md) and the three package runs of docs/benchmarks/repro.md. A submitted result agrees when every number is inside the tolerance the result file states (hash 3%, probe 10%, vectors and fingerprint exact). 'match' is a case-insensitive substring of the card name the workers report. The cpu gate is the chain's rule (under 10 ms per 32-lane warp on any core).",
|
|
"cpu_verify_ms_per_warp_max": 10,
|
|
"cards": [
|
|
{
|
|
"match": "RTX 5090",
|
|
"mhs": null,
|
|
"sweep_mhs": {
|
|
"4": null,
|
|
"64": null,
|
|
"256": null
|
|
},
|
|
"chase_gloads_1024": null,
|
|
"stream_gbps_1024": null,
|
|
"fingerprint": {
|
|
"24": "25f96e7dce90bd4e"
|
|
},
|
|
"source": "pending: the 5090-only re-run on PC 2 (the 6 October run was beside the live miner: 62.4 MH/s, not the card's figure; its fingerprint stands); the bench log's 5090 numbers before the package are 139.7 MH/s with the card to itself on a version 2 pack (4 October 2026, miner performance: variant racing) and 124.2 MH/s mining in the app"
|
|
},
|
|
{
|
|
"match": "Apple M5 Max",
|
|
"mhs": 27.674,
|
|
"sweep_mhs": {
|
|
"4": 275.247,
|
|
"64": 103.227,
|
|
"256": 56.625
|
|
},
|
|
"chase_gloads_1024": 3.401,
|
|
"stream_gbps_1024": 546.3,
|
|
"fingerprint": {
|
|
"24": "25f96e7dce90bd4e"
|
|
},
|
|
"source": "docs/benchmarks/repro-2026-10-06/mac (package v0.1.0-repro, Metal, 5 October 2026, run 2fdd7365db4a5deb, load average 6.19 at start); the bench log before the package: 27.9 MH/s through Apple OpenCL on the genesis pack (4 October 2026, proto-opencl README) and 26.7 MH/s mining through Metal"
|
|
},
|
|
{
|
|
"match": "gfx1201",
|
|
"mhs": null,
|
|
"sweep_mhs": {
|
|
"4": null,
|
|
"64": null,
|
|
"256": null
|
|
},
|
|
"chase_gloads_1024": null,
|
|
"stream_gbps_1024": null,
|
|
"fingerprint": {},
|
|
"source": "pending: the PC 1 package run; the bench log's 9070 XT numbers before the package are 18.0 to 18.1 MH/s in the bench and 2.42 to 2.68 G random reads/s at 1 GiB (5 October 2026, the 9070 XT on the eGPU)"
|
|
},
|
|
{
|
|
"match": "gfx1036",
|
|
"mhs": 3.312,
|
|
"sweep_mhs": {
|
|
"4": 3.722,
|
|
"64": 3.383,
|
|
"256": 3.32
|
|
},
|
|
"chase_gloads_1024": 0.458,
|
|
"stream_gbps_1024": 63.5,
|
|
"fingerprint": {
|
|
"24": "25f96e7dce90bd4e"
|
|
},
|
|
"source": "run-repro-pc2-20261006 (package v0.1.0-repro, AMD OpenCL, 6 October 2026, docs/benchmarks/repro-2026-10-06/pc2); the bench log before the package: 3.3 MH/s mining on the live devnet (4 October)"
|
|
}
|
|
]
|
|
}
|