igneum-network/igneum
1
0
Fork 0
Code Issues Pull requests Releases Activity

Bench log: the Arc B580 on PC 2's Razer Core X (10.98 MH/s, the link does not bind the hash; the dataset build 2.1x slower); the bench ROW line reads the adapter's driver instead of a fixed string

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
igneum-labs 2026-10-07 16:25:20 +00:00
parent 50e5d59183
commit 99f1b60f5d
2 changed files with 5 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
docs/bench-log.md
View file

@ -2683,3 +2683,5 @@ Found on the way: the Intel installer's bootstrapper wraps the inner installer's
Consequences per tier, measured: a B580 owner (12 GB, Windows) mines at 11.0 MH/s once the shipped worker carries 26e135a3 (0.3.20): about 2,450 blocks a day at the 388 MH/s the devnet showed at 11:19Z (one every 35 s, about 77,600 IGN a day at 31.69 IGN a block; approximate, the rate moves), at 150 to 190 W (approximate, no reading) about 0.06 to 0.07 MH/W, under the 9070 XT per watt (0.09) and about even per pound at a third of its price; mine only (the prover is CUDA only; the 12 GB are not the limit). Until 0.3.20 the card sits off on PC 1 (persisted) and the branch's app row says why. 8 GB and 16 GB Intel rows owed; Linux and HiveOS take the same worker and the same rewrite on `intel-opencl-icd`. The AMD and NVIDIA paths see no change (`proto-opencl/test_intel_rotr.c`, in the gate).
**PC 2, the same card in a Razer Core X V2 over Thunderbolt (7 October 2026, 16:14 to 16:16Z; job `run-ia-pc2-arc-bench-20261007-b`, the 0.3.20 worker sha256 af53f194..., the Arc alone through `--cards-off`, the 5090 and 4070 mining):** self-test PASS 96 of 96 on every pack, fingerprints 867dbc45cfb36b4d (the amended class v4 pack) and 90f794dd556f7a3b (the control), rates 10.982 MH/s class v4, 10.998 control, 10.989 live pack, 10.784 local exchange; the memprobe 1.403 G loads/s at 1 GiB (ceiling 10.96): within 0.3 percent of the same card in PC 1's PCIe slot (11.011, 1.41 G), so the eGPU link does not bind the hash (the dataset and cache live on the card; a job crosses the link with 32 bytes up and the hits down). What the link does cost: the cache and dataset build 192 + 683 ms against 107 + 320 ms in the slot (2.1x, the one-off upload per epoch), and the device's PnP properties read `CurrentLinkWidth 1, CurrentLinkSpeed 1 (2.5 GT/s)` with the same maxima, before and after the restart: either the Thunderbolt tunnel's downstream port reports x1 gen 1 to the OS or the property does not describe the tunnel; the 2.1x on a 1.25 GiB upload says about 1 GB/s effective, a Thunderbolt 3 or 4 PCIe tunnel's figure, not x1 gen 1's 250 MB/s; the reading is recorded as the property gave it, approximate. Driver: run 1 of the PC 2 driver job (exit 1014) and the job's own restart took the card from Code 12 on Windows' inbox 32.0.101.6733 to Code 0 on the same inbox driver with the Intel OpenCL platform live; run c binds 32.0.101.9034 (as on PC 1). The 0.3.20 app on PC 2 lists the card as `intel:Intel(R) Arc(TM) B580 Graphics`, enabled, 8 identities.

4
tools/intel-arc/pc1-arc-bench.ps1
View file

@ -180,7 +180,9 @@ $hoursBlock = if ($blocksDay -gt 0) { [math]::Round(24 / $blocksDay, 1) } else {
$ignDay = [math]::Round($blocksDay * $subsidyIgn, 1)
$shadowCost = if ($v4 -and $v3 -and $v3.mhs -gt 0) { [math]::Round((1 - $v4.mhs / $v3.mhs) * 100, 1) } else { 'owed' }
$share = if ($ceilingMhs -gt 0 -and $rate -gt 0) { [math]::Round($rate / $ceilingMhs, 2) } else { 'owed' }
"RESULT ROW card=`"Intel Arc B580`" driver=32.0.101.9034 device=$dev global_mib=$globalMib class_v4_mhs=$(if ($v4) { $v4.mhs } else { 'failed' }) class_v3_mhs=$(if ($v3) { $v3.mhs } else { 'failed' }) live_mhs=$(if ($live) { $live.mhs } else { 'failed' }) local_exchange_mhs=$(if ($v4local) { $v4local.mhs } else { 'failed' }) ceiling_mhs=$ceilingMhs latency_share=$share shadow_cost_pct=$shadowCost watts=not-exposed mhw=not-exposed network_mhs=$networkMhs blocks_per_day=$blocksDay hours_per_block=$hoursBlock ign_per_day=$ignDay card_alone=$cardAlone"
$drvNow = ''
try { $drvNow = (Get-CimInstance Win32_VideoController -ErrorAction Stop | Where-Object { $_.Name -match 'Intel' } | Select-Object -First 1).DriverVersion } catch { }
"RESULT ROW card=`"Intel Arc B580`" driver=$drvNow device=$dev global_mib=$globalMib class_v4_mhs=$(if ($v4) { $v4.mhs } else { 'failed' }) class_v3_mhs=$(if ($v3) { $v3.mhs } else { 'failed' }) live_mhs=$(if ($live) { $live.mhs } else { 'failed' }) local_exchange_mhs=$(if ($v4local) { $v4local.mhs } else { 'failed' }) ceiling_mhs=$ceilingMhs latency_share=$share shadow_cost_pct=$shadowCost watts=not-exposed mhw=not-exposed network_mhs=$networkMhs blocks_per_day=$blocksDay hours_per_block=$hoursBlock ign_per_day=$ignDay card_alone=$cardAlone"
"RESULT PICKER ['Intel Arc B580', $rate] (site/yourcard.js CARDS entry; measured, so it may go on the site)"
$prove = 'mine only: the prover is CUDA only, and even a 12 GB NVIDIA card cannot mine and prove at once on this build'
"RESULT CONSEQUENCES a B580 owner (12 GB, Windows): $rate MH/s on the OpenCL worker, $blocksDay blocks a day at today's $networkMhs MH/s network (one every $hoursBlock h, about $ignDay IGN a day at $subsidyIgn IGN a block, approximate: the network rate moves); hash per watt not measurable yet (no power reading through OpenCL; IGCL reader next, 2 to 3 hours); $prove; 8 GB and 16 GB Intel rows owed (no card in hand); Linux and HiveOS take the same worker on intel-opencl-icd (owed a line)"