The record of 12 October 2026. This page says what "open hardware" is on Igneum today, item by item, with the record behind each and the public path where one exists, and it says plainly what is not published yet. Nothing on it is a promise; the dates are the record's.
1. The reference rig, as built
An Octominer X8 ULTRA PLUS: eight full-size x16 slots at x1 electrical, a Pentium G4400T, 8 GB of memory, a 60 GB SATA SSD, four 750 W server supplies in the bays, four chassis fans; a commodity mining chassis anyone buys. The cards seated on it, by the rig's own bus read in the machines feed (https://build.igneum.network/feed/machines.json, the chassis entry, read 2026-10-12 18:24Z): 02:00.0 NVIDIA GeForce RTX 5060 Ti (not bound on the rig's driver tonight); 03:00.0 NVIDIA GeForce RTX 3060; 04:00.0 AMD Radeon RX 580; 05:00.0 NVIDIA GeForce RTX 2060 KO; 08:00.0 AMD Radeon RX 7600; 0B:00.0 AMD Radeon RX 5700 XT; 0E:00.0 AMD Radeon RX 6600; 11:00.0 AMD Radeon RX 9060 XT. One card a slot, the 8-pin leads from the bay supplies to every card with a socket, one Ethernet cable to the house router; the feed is the list of record, this sentence its copy at that minute. The host runs an unattended Debian 13 image with the bench agent baked in; the only hands are the stick, the cards and one boot. The record: the rig's record (the recipe, the earliest boot, the lanes, the cause of the earliest night's resets) and the BENCH-JOB-FORMS file section 10 (the image).
| Part | What | Public path |
|---|---|---|
| the chassis | Octominer X8 ULTRA PLUS (board "INTEL B85 ULTRA", BIOS 5.011): 8 slots, 4 bay supplies of 750 W, 4 chassis fans | a commodity part; no file |
| the host | Pentium G4400T, 8 GB, 60 GB SATA SSD, rear VGA, USB type A only | a commodity part; no file |
| the image | Debian 13 unattended, the bench agent, the earliest-boot unit, the fan unit; built on a build box by the build-iso file | https://build.igneum.network/rig/igneum-rig-debian13.iso with SHA256SUMS (the iso's sha256 33d3eccc...38f0) and image.txt (the netinst's pinned sha, the agent's version and sha, the repository sha of the build) |
| the bench agent | one Python file, the standard library, a signed job queue, the machine rows | https://build.igneum.network/releases/bench/ (0.2.40 at this writing; image.txt names the version baked in) |
| the workers | CUDA, OpenCL (and Metal on Apple silicon) under the workers directory, the CPU reference under the ref directory, the miner under the miner directory | not published as source (section 5); fetched by the agent by sha from the artefact store for each job; Ember ships as signed binaries at https://build.igneum.network/releases/ember/ |
| the drivers | NVIDIA 550.163 and AMD OpenCL (rusticl, Mesa 26.1) fetched on the earliest boot from Debian's mirrors; Intel compute not served on the rig | Debian's own |
What the earliest night taught, kept in the record: the Octominer's fan controller (a small V-USB board on the internal USB header) carries a hardware watchdog that resets the host 25 minutes after its power-up unless the vendor's software feeds it; the rig's rule is that its ribbon stays unplugged, the chassis fans run at full from their headers, and the thermal guard reads the cards, not the fans (the rig's record section 9, the cause closed on 12 October 2026).
2. The power and thermal rules
- Power at the wall: one smart plug per machine and per external card housing, registered once in one file, read by the agent through an in-house relay that polls each plug once a second; idle and load rows written beside every cell that asks the wall grade (the METERS file). A rented host has no meter, so its rows are telemetry grade for ever; only an owned machine with a plug at the wall produces a wall-grade row (the WALL-METER-PROTOCOL file).
- What a wall-grade row is: accepted work per wall-kWh as a within-card paired comparison, five paired runs of thirty minutes each, the candidate and its control back to back on the same card, the ratio bootstrapped over the pairs; never a cross-algorithm rate, never a figure from a card the team does not own.
- Thermal: the cards' own fans and temperatures through the driver (NVML, amdgpu's hwmon); the chassis fans on the reference rig at full from their headers (the board exposes no controllable fan chip to Linux; the rig's record 3a and 9); the thermal guard in the agent reads the cards.
- Tuning: every point of the tuning round is pre-registered before the earliest decisive run, every decisive cell carries a paired 4 GiB row in the same hour, and a card's power sweep gives the energy per accepted work with bounds (the TUNING-R0A file); the per-family recipes (NVIDIA, AMD GCN and RDNA, Apple) are the rows below and the two family notes the AMD-COHORT file and the AMD-WAVE-PATH file.
3. The measured rows
The cohort of record (the cohort record file, updated 2026-10-12T17:56:00Z): one row per card and boundary at the gate point of the launch document unless the form says otherwise. The grades are the record's own words: measured, estimated, pending, screening, model; the boundary says whether the row was read from the card's telemetry or from a meter at the wall. A rented pod's row is telemetry grade by rule.
| Vendor | Card | Rate (MH/s) | Energy (uJ per hash) | Instrument | Grade | State | Note |
|---|---|---|---|---|---|---|---|
| NVIDIA | RTX 5090, rented pod, 500 W cap | 40.6 | 12.20 | Telemetry | measured | The energy row of record; stable repeats and the production parity row retained as requirements | The programme's energy 16.5 percent lower at the 450 W knee |
| NVIDIA | RTX 5090 Founders Edition (bench PC A), owned, 450 W cap | 37.0 | 12.06 | Telemetry | measured | The FE's cap-point row; the free-clock 575 W rows are the FE's rows of record beside it (the coordinator, 11 October) | The wall row waits on the wired WireView Pro II (the FE takes no plug-in meter; due 5 November 2026); the machine plug is the whole machine's boundary |
| NVIDIA | RTX 5090 Founders Edition (bench PC B), owned, 575 W cap, free clock | 13.44 | Telemetry | measured | The FE's free-clock row of record; the pods' 500 W rows and the locked twins are the energy rows of record (the independent reviewer) | Overlap 1.151, m_blend 0.210; the walk 3.25 to 3.28 uJ at 172 to 173 MH/s | |
| NVIDIA | RTX 4090 (bench PC B, enclosure 2R), owned, 480 W cap, uncapped | 51.4 | 8.38 | Wall-differential | measured | Three 180 s windows; 9.24 uJ by the sensor at 478.6 W NVML | The walk 95.8 MH/s at 3.65 uJ the cell gb-enc__rtx4090-24g__wall-2 |
| NVIDIA | RTX 4090 (bench PC B, enclosure 2R), the 380 W knee, owned, 380 W cap | 46.3 | 7.05 | Wall-differential | measured | The most energy-efficient row of any card at a wall; 8.09 uJ by the sensor at 379 W NVML | The knee buys 16 percent of energy per hash for 10 percent of rate; the candidate comparator for the specialist ratio is fixed in the record the cell gb-enc__rtx4090-24g__wall-2 |
| NVIDIA | RTX 4090, rented pod, 400 W cap | 46.64 | 8.531 | Telemetry | screening | The rented screening row; the owned card's wall rows above are the rows of record | the cell igneumhash-rung-k16-screen__rtx4090-24g-b__screen |
| NVIDIA | RTX 5080 (bench PC A, enclosure 1L), owned, 360 W cap | 22.9 | 13.54 | Wall-differential | measured | Three windows; the card's own telemetry column empty in the OpenCL form | The walk 103.0 MH/s at 3.01 uJ |
| NVIDIA | CMP 170HX (the rig, as shipped), owned, 250 W cap | 17.1 | 11.30 | Wall-differential | measured | Three windows 17.13, 17.12, 17.13 MH/s at 205.9, 203.6, 205.4 W NVML; 190 to 196 W attributable at the rig's plug; 11.9 uJ by the sensor; the launch document's five keyed sets exact | The shipped card on the stock driver at its 1,410 MHz clock ceiling (not its cap); the wide FP op's form 10.5 percent faster than the integer form; an unlocked row only on the project's go; the walk 55.2 MH/s at 3.7 uJ the cell gb-rig-row__cmp170hx-8g__wall-5 |
| NVIDIA | RTX 2060 6 GB (the rig, the owned floor-class card), owned, 172 W cap | 7.9 | 17.60 | Wall-differential | measured | The floor-class row of record for G-32: three uncapped 90 s windows at 171 W board (21.4 uJ by the sensor); at the 138 W cap 7.33 MH/s at 15.5 uJ wall-differential (the card's own knee) | The stock-limit row of 12 October 17:20Z: 7.9 MH/s at 171.5 W NVML = 21.7 uJ per hash telemetry on three windows (69 C max), the W128 walk 15.9 MH/s at 153.8 W = 9.7 uJ; the same rate as the morning's row at the card's full 172 W limit against the cap step's 138 W, so 1.23x the energy for the cap alone; the plug read every parallel lane, no wall figure from this run; the morning's wall-differential row stays the row of record and this is the stock-limit telemetry row beside it. The card is a STANDARD RTX 2060 6 GB (TU106, the bus string 'GeForce RTX 2060 Rev. A'), not the KO edition the order and the morning's rows named; the rows stored under rtx2060-ko-6g are this card under the old name; new rows are named rtx2060-6g. the rig's earlier freezes were the RX 6600's and the board's, not this card's; the stepped row ran with the host alive throughout the cell gb-rig-step__rtx2060-ko-6g__cap-1 |
| AMD | RX 9070 XT (bench PC A, enclosure 2R), the card of record, owned | 15.5 | 13.34 | Wall-differential | measured | AMD family qualification passed; the wall confirmation IS this row (13.43, 13.27, 13.31 uJ); the telemetry pair of record 13.92 uJ (14.0 to 14.06 by ADLX on the same windows: the sensor reads 11.7 W above the plug, so the telemetry rows are conservative on this card) | 1.14x the 5090 pod by telemetry (13.92 against 12.2); the walk 33.3 MH/s at 6.23 uJ the cell gb-enc__rx9070xt-16g__wall-2 |
| AMD | RX 9060 XT, the corroboration, owned | 14.24 | Telemetry | measured | 14.0 to 14.3 uJ across two windows, 1.16x the 5090; exact and spill-free on every class | Independent RDNA 4 corroboration of the card of record; no wall row; its wall row on the rig is queued tonight (gb-rig-row__rx9060xt__wall-2, the generic path); its the rig generic-path row of 12 October (gb-rig-row__rx9060xt__wall-3, sealed ok 17:24:36Z, every class exact including the wave form's): RATE ONLY, the generic form 2.31, 2.31, 2.28 MH/s on three 180 s windows, the W128 walk 7.07 MH/s, the wave window's output empty; the plug's figure VOID (the chassis at 1,017 to 1,037 W through the windows, another lane's draw on the one plug) and no hwmon on the rusticl worker; its bench PC B telemetry row stays the row of record; its R0C row runs on the rig now | |
| AMD | Instinct MI300X, rented pod, 750 W cap | 47.93 | 11.73 | Telemetry | screening | The data-centre AMD part at the hwmon cap; 0.95x the 5090's energy | |
| Apple | M6 mini (16 GB), owned | 1.97 | Estimate | estimated | The rate measured (1.97 and 1.96 MH/s, double the earlier Metal kernel); thirty leak-free warm transitions at 2 GiB; the energy at the gate point estimated from the package band | The mini's plug is LIVE (a Tapo P110 on its own mains lead since 11 October ) and carries the mini's walk-form wall row of record (or-walldec-5: 7.74 MH/s at 37.8 W whole machine, 4.89 uJ per hash whole-mini, 4.49 above its 3.1 W idle); the gate-point wall row is one queued job on that plug and lands the hour it runs the cell or-formb-v8s-1 | |
| Apple | M5 Max (64 GB, the 14-inch laptop), owned, 100 W cap, High Power | 5.4 | Estimate | estimated | The rate measured (5.43 and 5.31 MH/s against 2.8 on the earlier kernel); the energy estimated | No plug on the laptop: the fourth Tapo went to the rig, and a laptop's wall row needs the battery kept full under the 96 W charging ceiling of the 14-inch; its wall row waits on a plug placed on it, no date; its own kernel work, soak and full-battery repeat open the cell mb-formb-v8s-1 | |
| Intel | Arc B580 (bench PC B, enclosure 2R), owned | Telemetry | pending | Exact on every row (SIMD16 pinned); no sealed gate-point energy row yet | the cell gb-intel-b580__tune-3 | ||
| NVIDIA | RTX 3060 12 GB, owned, 170 W cap | 8.5 | 19.90 | Telemetry | measured | Sealed ok 17:18:30Z on the rig's lane (the CUDA worker 1e65533d, exactness 96 of 96 on both sets): 8.52, 8.52, 8.53 MH/s at 169.7 W NVML = 19.9 uJ per hash on all three windows (max 51 C); the W128 walk 28.2 MH/s at 169.5 W = 6.0 uJ; TELEMETRY grade: the plug read every parallel lane (the per-2 s NVML lines in the cell for a subtraction) | GA106, Ampere, 12 GB; the parallel form (agent 0.3.49) stated; its earliest run thermal-stopped at 16:51:48Z two seconds beside the 2060's, both restarted the cell gb-rig-row__rtx3060-12g__wall-2 |
| AMD | RX 580 (the rig), owned, the generic OpenCL path, untuned | Wall-differential | pending | Queued on the rig at 17:14 UK (on the Linux OpenCL worker 1135c79d, the split-buffer build), 12 October, back to back (the 580, the 7600, the 5700 XT, the 9060 XT's wall row, the 6600 last); the row lands here the hour it seals | Polaris, GCN 4 (gfx803); the chassis fans off, the card on its own fans with the thermal guard at 80 C edge; a freeze or the guard ends the row with its row; re-issued holding on the device id 1002:67df (the agent's Polaris string lists the family bracket), so it runs after the 3060's queued job the cell gb-rig-row__rx580-8g__wall-3 | ||
| AMD | RX 7600 (the rig), owned, the generic OpenCL path (Mesa rusticl on radeonsi), untuned | 1.01 | None | measured | RATE ONLY: sealed ok 17:17:31Z on the rig (worker 1135c79d; fp-v8, the v4 and v8 classes and the wave form's classes all exact): the generic local-exchange form 0.96, 1.02, 1.06 MH/s over three 180 s windows; the family wave kernel (form1-wave-v8s) 3.82 MH/s on its one window; the W128 walk 4.26 MH/s. ENERGY NONE: the Linux OpenCL worker's stat lines carry no hwmon watts, and the rig's plug read every parallel lane (the chassis at 408 to 1,042 W through the windows), so no wall figure; the card's row of record on energy waits on a telemetry line or a serial wall run | Navi 33, RDNA 3; the chassis fans on for the window, the card on its own fans with the thermal guard (72 C edge, 93 C hotspot peak, reliability's read); the parallel form (agent 0.3.49, per-card lanes under a 1,500 W plug guard, no pause) stated; the family kernel reads 3.8x the generic path on this card, so the tuned row is the one to serve when it is cut the cell gb-rig-row__rx7600-8g__wall-3 | |
| AMD | RX 5700 XT (the rig), owned, the generic OpenCL path, untuned | Telemetry | pending | Queued on the rig at 17:14 UK (on the Linux OpenCL worker 1135c79d, the split-buffer build), 12 October, back to back (the 580, the 7600, the 5700 XT, the 9060 XT's wall row, the 6600 last); the row lands here the hour it seals | Navi 10, RDNA 1; the chassis fans off, the card on its own fans with the thermal guard at 80 C edge; this row's energy is TELEMETRY grade (the card's hwmon), the overlap stated; the wall-differential form re-runs serially at the end of tonight's queue only if the telemetry figure disagrees with the 9070 XT family's by more than the record's tolerance, else the telemetry figure stands as the row's grade. The 5700 XT's row follows the 7600's in the serial order and overlaps the 3060's window in part; the windows inside it are telemetry grade, named per window in the readout. the cell gb-rig-row__rx5700xt-8g__wall-2 | ||
| AMD | RX 6600 (the rig), owned, the generic OpenCL path (Mesa rusticl on radeonsi), untuned | 0.94 | None | measured | RATE ONLY: sealed ok 17:26:56Z on the rig's lane (worker 1135c79d; fp-v8, the v4 and v8 classes and the wave form's classes all exact; NO FREEZE under the thermal guard, the card on its own fans): the generic form 0.95, 0.95, 0.93 MH/s over three 180 s windows; the family wave kernel 4.05 MH/s on its window; the W128 walk 3.68 MH/s. ENERGY NONE: no hwmon watts on the Linux OpenCL worker's stat lines and the plug carried every parallel lane (the chassis at 903 to 931 W through the windows) | Navi 23, RDNA 2; the parallel form (agent 0.3.49, per-card lanes under the plug guard) stated; the five freezes of 11 October did not recur under the guard tonight, so the slot or the feed under that night's load form is still the open row, not this card the cell gb-rig-row__rx6600-8g__wall-7 |
Not in the table, no cohort figure: RTX 5090 Founders Edition, the chain's hour (Ember on devnet, the share lane, blocks 555 to 673); RTX 5060 8 GB (no gate-point row yet); RTX 5060 Ti 16 GB (the W128 walk only (screening)); RTX A5000 24 GB (no gate-point row yet); RX 580 8 GB (the current form (the blended version 8 twin b2-x16-f4128) on the GENERIC PATH (Mesa rusticl, the generic OpenCL kernel)).
Four rows carry no gate-point figure and say so in their form: the RTX 5090's hour on the chain's share lane (a measured rate, not the gate point), the RTX 5060 Ti's screening walk, and two pending rows with no form yet (the RTX 5060 8 GB and the RTX A5000); a served table that prints only gate-point figures leaves them under it, named.
Every row's key names the cell it came from; the cells are under the igneumhash directory by name.
4. The public paths today
| What | Path | How it is checked |
|---|---|---|
| the rig image | https://build.igneum.network/rig/ (igneum-rig-debian13.iso, SHA256SUMS, image.txt) | sha256sum -c SHA256SUMS |
| the bench agent | https://build.igneum.network/releases/bench/ | the release's SHA256SUMS and manifest |
| Ember, the miner | https://build.igneum.network/releases/ember/ (0.2.1 at this writing: macOS, Windows; SHA256SUMS.sig and MANIFEST.json.sig beside) | SHA256SUMS.sig against https://build.igneum.network/keys/allowed_signers (namespace igneum-release) |
| the fleet's machines and cards | https://build.igneum.network/feed/fleet.json | read |
| the chain's node, wallet, pool and developer kit | https://build.igneum.network/releases/ | the same sums and signatures |
| the release key | https://build.igneum.network/keys/ | the fingerprint on the keys page |
5. What is published under what, and what is not published yet
- Published: the rig recipe and its image (section 1); the agent as a release; Ember, the node, the wallet, the pool and the developer kit as signed releases with their builders' manifests; every measured row above with its grade; the fleet feed.
- Not published yet: the worker sources (the workers directory, the ref directory, the miner directory), the agent's source, the image's build script and the rest of the repository. The repository is private; nothing serves its source today. A public mirror of the sources at a pinned sha is a decision of the project's, and this page names the date it is taken.
- The licence: none is chosen yet. The repository carries no licence file; the third parties it vendors keep their own (OpenZeppelin, the Core-Geth lineage, forge-std, the account-abstraction contracts, each pinned with its licence; the explorer runs under Blockscout's own licence; a reference hash implementation without a licence file is measured against and never redistributed). The licence the sources are published under is the project's decision, and this page names it the day it is made. Until both decisions are taken, "open hardware" on Igneum means an open rig recipe, an open agent release, open measured rows and open binaries, with closed sources; the page says so rather than say more.