diff --git a/docs/evidence.md b/docs/evidence.md index 4c63f5c02..49d9cb1d3 100644 --- a/docs/evidence.md +++ b/docs/evidence.md @@ -32,7 +32,8 @@ Versions in the table: "the release manifest" is `site/release-manifest.json`, s | 3 | The 64-register window per lane costs a GPU under 1 percent of rate at stock, and at most 5 percent per load with the liveness chain | The litepaper (class v6); `docs/plans/igneum-2.0.md` D1 (the placed 64-register rows) | tested by the team | `docs/analysis/class-v6/connected-state.md` section 4; `docs/design/class-v6-rotating-family.md` section 10.0e | The class v5 nvcc harness and the kit worker, both packs on the same card minutes apart, 250 batches of 2^24, nvidia-smi at 1 Hz, vectors PASS on every row (`connected-state.md` section 4); the per-load rows of the full chain against the base (`class-v6-rotating-family.md` 10.0e) | 8 October 2026, rented RTX 5090 (575 W cap) and RTX 4090 (450 W cap) at stock: energy per hash +0.6 percent on the RTX 5090 and -0.9 percent on the RTX 4090, inside the run-to-run noise; under 1 percent of rate; 80 to 87 registers per thread, no spill (all measured). Per load: RTX 5090 16.7 nJ base, 17.6 nJ full chain; RTX 4090 26.0 nJ, 27.0 nJ (measured). The lock row on the project's own rigs is owed | none yet | | 4 | Reorganising the same work around live state (the connected-state variant, experiment D2(a)) does not reduce a specialised chip's edge: KILL as a class | `docs/plans/igneum-2.0.md` D2(a); this page | tested by the team (a published failure) | `docs/analysis/class-v6/connected-state.md` (the verdict, section 6) | The census, liveness and GPU rows in `connected-state.md` sections 2 to 4; the chip side priced on the drawn program by synthesis (a model, never a lower bound) | 8 October 2026, verdict 17:25 UK: the window is necessary (63 of 64 registers live at every address, measured) but only its width reaches the chip, +1.2 pJ per lane-op at N5 (synthesised); the window moves the chip's edge 1.10x node for node against a 1.25x gate (modelled); the GPU side +0.6 percent energy per hash on the RTX 5090, -0.9 percent on the RTX 4090 at stock (measured). Rearranging the dependency graph of the same operations moves neither side | none yet | | 5 | "A GPU-secured network for Ethereum-compatible applications and verifiable computation." served on every page | Every page | designed (served) | `docs/plans/igneum-2.0.md` (the objective: the positioning line) | `node tools/ci/ledger-text-check.mjs`: the sentence pinned (R0) on the home page, the litepaper and this page | 8 October 2026: on this page; the home page and the litepaper carry it as their 2.0 text lands (designed) | none yet | -| 6 | The chip claim as served: "Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes." Under it the three statements, separate: energy (the modelled bracket about 2.3x to 3.3x a node ahead and 2.0x to 2.9x node for node, approximate and provisional until the placed gated core rows land), economic and response capability, rotation an optional improvement. Class v5 derives the dataset from chain state; whether that excludes a specialised design is under evaluation (Deliverable 3), since a design that tracks state is not excluded by staleness. The energy ratio is not the pass criterion: the coexistence model ([docs/analysis/class-v6/coexistence-model.md](https://git.igneum.network/igneum-network/igneum/src/branch/master/docs/analysis/class-v6/coexistence-model.md)) is, and its first run's result is served with its conditions | The litepaper (the chip model) | designed (the bracket modelled; the GPU side tested by the team) | `docs/design/class-v6-rotating-family.md` section 10 (10.0h to 10.0n, 8 October 2026); `docs/plans/igneum-2.0.md` D3 and D4 | the scoring rule in the close (the minimum over workloads of the maximum over free adversarial designs of the GPU's joules per hash over the adversary's, under the 10 percent GPU-cost budget at the lock, the verifier limit, cross-vendor correctness and hardware accessibility); the placed rows are D3's | the GPU side measured: the RTX 5080 at its 1,100 MHz lock 2.06 microjoules per hash and the RTX 5090 at its 1,300 MHz lock 2.33 (8 October 2026, the project's own rigs and rented pods); the chip side synthesised and claimed, its placed gated row pending | none yet | +| 6 | The chip claim as served: "Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes." Under it the three statements, separate: energy (the modelled range 2.5x to 3.0x a node ahead and 2.1x to 2.6x node for node, the placed row to narrow it), economic and response capability, rotation an optional improvement. Class v5 derives the dataset from chain state; whether that excludes a specialised design is under evaluation (Deliverable 3), since a design that tracks state is not excluded by staleness. The energy ratio is not the pass criterion: the coexistence model ([docs/analysis/class-v6/coexistence-model.md](https://git.igneum.network/igneum-network/igneum/src/branch/master/docs/analysis/class-v6/coexistence-model.md)) is, and its first run's result is served with its conditions: The test is whether a specialised supplier can earn a normal return while ordinary GPUs stay close enough in total cost, obtainable and useful outside mining that new entrants still compete. On today's modelled rows that holds for the chip anyone can build, a stored-dataset board on commodity DRAM, at a productive life of one to three years: its cost per accepted unit of work sits within the range of the best GPU owner and entrant, and a fleet of it holds a minority of the network with GPU entrants still setting the price. It fails for the SRAM-store die once that die exists with its development paid: no electricity price keeps a GPU owner level with it, and a modest fleet takes most of the network. What holds the die is the investment decision, not the hash; every chip figure here is modelled, not measured, and the chip's hardware cost per unit of work is approximate within 2x. | The litepaper (the chip model) | designed (the bracket modelled; the GPU side tested by the team) | `docs/design/class-v6-rotating-family.md` section 10 (10.0h to 10.0n, 8 October 2026); `docs/plans/igneum-2.0.md` D3 and D4 | the scoring rule in the close (the minimum over workloads of the maximum over free adversarial designs of the GPU's joules per hash over the adversary's, under the 10 percent GPU-cost budget at the lock, the verifier limit, cross-vendor correctness and hardware accessibility); the placed rows are D3's | the GPU side measured: the RTX 5080 at its 1,100 MHz lock 2.06 microjoules per hash and the RTX 5090 at its 1,300 MHz lock 2.33 (8 October 2026, the project's own rigs and rented pods); the chip side synthesised and claimed, its placed gated row pending | none yet | +| 7 | Mining and proving together on one card needs a 16 GB card at the 5.5 GiB dataset floor: the miner holds about 6.1 GiB and a compressed shard proof peaks at about 7.5 GiB, so 8 GB and 12 GB cards time-share (the app pauses the miner for the proof). NVIDIA proves; AMD and Apple mine | The litepaper (Proving, vs RandomX); the miner page | tested by the team | `docs/analysis/class-v6/coexist-rows.md` (the 5.5 GiB ds55 miner beside igneum-prove-host-0317, compressed at threshold 2^26, the served sm_86 and sm_89 floors) | the RESULT rows in that file, verbatim from the runs | rented RTX 3060 12 GB: 26.82 MH/s at 117.4 W with 6,129 MiB resident; the compressed shard 13.2 s, peak 7,525 MiB (6,129 + 7,525 = 13,654 MiB against 12,288); rented RTX 4060 8 GB: 18.84 MH/s, 6,116 MiB resident; the shard 8.2 s, peak 7,532 MiB; 8 October 2026 | none yet | ## Count by status @@ -41,11 +42,11 @@ Versions in the table: "the release manifest" is `site/release-manifest.json`, s | designed | 3 (rows 1, 5, 6) | | implemented | 1 (row 2) | | activated | 0 | -| tested by the team | 2 (rows 3, 4) | +| tested by the team | 3 (rows 3, 4, 7) | | reproduced externally | 0 | | reviewed independently | 0 | -6 rows. The rendered page is `site/evidence.html` (served at /evidence), generated from this file by `site/build.mjs`; the text is judgement, so this file is edited by hand and the page follows. +7 rows. The rendered page is `site/evidence.html` (served at /evidence), generated from this file by `site/build.mjs`; the text is judgement, so this file is edited by hand and the page follows. ## What would move a row diff --git a/site/app.html b/site/app.html index 78d8cb5bc..9fe588e9d 100644 --- a/site/app.html +++ b/site/app.html @@ -360,14 +360,14 @@
Prove
The Prove page of the current build: the shard card, your card proved shard 3 of block 160,390 for 1.15 IGN; the Prove on this machine switch on, one sentence for the Apple M5 Max, which proves on the CPU slowly; one line of counts, and Details The Prove page of the current build in light mode: the shard card and the Prove on this machine switch -
The current build, rendered from the recorded state of the team’s Apple M5 Max, 7 October 2026. Apple silicon proves on the CPU, slowly; a 12 GB NVIDIA card is the line for proving beside the miner.
+
The current build, rendered from the recorded state of the team’s Apple M5 Max, 7 October 2026. Apple silicon proves on the CPU, slowly; a 16 GB NVIDIA card is the line for proving beside the miner at the 5.5 GiB dataset floor, and an 8 GB or 12 GB card time-shares with the miner paused (measured 8 October 2026).
05 · Prove

One switch. The same card proves.

Every block is turned into a short proof, in pieces called shards. Your cards prove the shards the chain assigns to them and earn IGN for each one.

diff --git a/site/claims.html b/site/claims.html index 509a43aee..3fa5454c4 100644 --- a/site/claims.html +++ b/site/claims.html @@ -247,7 +247,7 @@

Here are the limits, stated before anyone else states them.

diff --git a/site/litepaper.html b/site/litepaper.html index 28cd9a945..d7b5d4bc9 100644 --- a/site/litepaper.html +++ b/site/litepaper.html @@ -339,7 +339,7 @@ body.all .pager{display:none}

Abstract

-

A GPU-secured network for Ethereum-compatible applications and verifiable computation. Igneum is a proof-of-work chain built for graphics cards. AMD, Apple and NVIDIA cards mine; NVIDIA cards also prove its blocks with zero-knowledge proofs. Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes. Class v6 adopts the 64-register window and retains it across every rotation. Current modelling places the strongest specialised designs assessed against the GPU tier at about 2.3x to 3.3x energy-efficiency advantage a node ahead (2.0x to 2.9x on the GPU's own node), a bracket that is approximate and provisional until the placed gated core rows land. The long-program and select-tree proposals were rejected. Economic resistance depends on development cost, deployment economics and productive hardware lifetime; family transitions receive an obsolescence benefit only where a loss of competitiveness is demonstrated; programmable multi-epoch designs are included in the assessment. The chip model: every number labelled measured, modelled or claimed, the harness and the scoring rules beside it.

+

A GPU-secured network for Ethereum-compatible applications and verifiable computation. Igneum is a proof-of-work chain built for graphics cards. AMD, Apple and NVIDIA cards mine; NVIDIA cards also prove its blocks with zero-knowledge proofs. Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes. Class v6 adopts the 64-register window and retains it across every rotation. Current modelling estimates a 2.5x to 3.0x energy-efficiency advantage for the strongest specialised designs assessed against the GPU tier (2.1x to 2.6x on the GPU's own node). The long-program and select-tree proposals were rejected. Economic resistance depends on development cost, deployment economics and productive hardware lifetime; family transitions receive an obsolescence benefit only where a loss of competitiveness is demonstrated; programmable multi-epoch designs are included in the assessment. The chip model: every number labelled measured, modelled or claimed, the harness and the scoring rules beside it.

It runs the Ethereum virtual machine through revm, so contracts built for Ethereum deploy with familiar tools; the differences (block context, randomness, two-dimensional fees) are documented. Transactions are included in about one second, proven within about a minute at launch, and locked by miners within about two (designed targets). There is no premine, no pre-sale, no treasury taken from emission, no stake anywhere in consensus, and no dependence on any other chain. Mining is built to stay open to anyone with a GPU. The argument rests on the scoring rule's result against placed adversary designs and on the economics, reported separately as energy advantage, economic advantage and response capability. The hourly program is an optional improvement, not the defence. Writing new code, including an emergency fix to the proof system, is the one thing that takes a person, and it activates only on miner signalling.

1 / s
blocks, rising to 10 (designed)
@@ -464,12 +464,12 @@ body.all .pager{display:none}

What carries the chip resistance is the scoring rule's result on the placed adversary rows, plus the economics, reported separately in the chip model below. Three properties feed it. Rotation is an option, not the defence. A new program every hour, drawn from the chain, its memory pattern with it, and rules that change on a schedule fixed at launch. These schedule changes defeat a chip wired for one datapath and need no human fork, but against a programmable chip that stores the dataset every drawn parameter is firmware, so the security argument does not rest on them. What meets that chip is the latency-shadow work (class v4) and the price per joule (the chip and economy analysis of 6 October 2026, section 5.4; ledger M32). It waits on memory, not maths. Every hash is a chain of random reads into a table too big for a chip to carry. Measured (8 October 2026; lane D’s family harness at the acceptance rule’s own 2^20 sample over 4,900 drawn eras, and the chained-cache pass’s reading of the night before): every hash’s 128 dependent reads land across the whole dataset and the distinct-index floor holds at 0.995 on every accepted program; about half of epochs carry one load site whose address bit at the era’s stride rotation is biased, which prices about 1.6 percent of a hash’s reads to a chip storing half the dataset and nothing to a chip storing all of it; the next class folds the product’s low bits before the rotation, so no era lands a biased bit on an address bit. The wait is the same physics for everyone. Miners hold the switch. Spare defences are written into the rules, switched off. A miner signal turns one on, at the class-change threshold: miners signal three things at three thresholds, 60 percent of blue blocks over two weeks for a parameter genesis leaves open, 90 percent for an upgrade (new code), and 95 percent with a floor height for a class change. No fork.

The work that waits can grow. Class v4 adds a block of latency-shadow arithmetic to every hash, about 100,000 integer operations that run while the memory reads are in flight, so a chip that stores the whole dataset still has to pay for a core. That size sits on a ladder fixed at genesis, six rungs from about 100,000 to about 1,000,000 operations, and it moves one rung at a time only when 90 percent of blue blocks in each of seven consecutive days ask for it; it can never move two rungs inside a week and never past a rung the reference verifier cannot check under 10 ms with its sibling thread busy (measured on the build server, 6 October 2026: the first three rungs pass at 8.8, 8.9 and 9.2 ms, the fourth misses by 0.08 ms on a loaded box and stays out until a quiet re-measurement, the two doublings are out at 12.4 and 15.0 ms). What it buys against a chip is scored under the rule in the chip model below. What it costs, per rung, is measured too: the Apple tier gives up 3 points of rate at the first step and 6 more at the second, the RTX 5090 nothing until the second; so the miners who pay for a step are the ones who take it.

The chip model

-

Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes. Class v6 adopts the 64-register window and retains it across every rotation. Current modelling places the strongest specialised designs assessed against the GPU tier at about 2.3x to 3.3x energy-efficiency advantage a node ahead (2.0x to 2.9x on the GPU's own node), a bracket that is approximate and provisional until the placed gated core rows land. The long-program and select-tree proposals were rejected. Economic resistance depends on development cost, deployment economics and productive hardware lifetime; family transitions receive an obsolescence benefit only where a loss of competitiveness is demonstrated; programmable multi-epoch designs are included in the assessment.

-

The labels. The bracket is modelled and provisional: its floor is the clock-gated base core and its ceiling the first placed core, which came in 64 percent above synthesis (wires and the clock tree); the honest figure is the placed gated core’s and replaces the bracket when its row lands. The GPU side is measured: an RTX 5080 at its 1,100 MHz core lock, 2.06 microjoules per hash, and an RTX 5090 at its 1,300 MHz lock, 2.33 microjoules per hash, both under class v4, on the project’s own rigs and rented pods, 8 October 2026; the card’s cost of the window is measured too (a rented RTX 5090 and RTX 4090 at stock, 8 October 2026: within 5 percent per load with the liveness chain, no register spill). The chip side is synthesised and claimed: the clock-gated sequencer core with the 64-register window on ASAP7, scaled to N3 on the foundry’s headline factors (k about 0.37 at N3 and 0.51 node for node for the base core, the gated window adding about 0.13 of k against an adversary with a flop register file; the window’s liveness measured at 61 of 64 values necessary, its cost to the card measured under 5 percent); the window’s k is synthesis-derived and not a lower bound, and the multi-family adversary lane’s first core (its state in a macro) reads the window’s defence as close to nothing, a disagreement between two models that the placed rows settle. The chip’s memory is modelled: the GDDR7 board of the chip model. The placed gated figure is expected near 2.6x to 3.1x a node ahead and 2.3x to 2.7x node for node (approximate) and is served when its row lands.

+

Igneum remains competitive on accessible commodity GPUs even when specialised mining hardware is assumed to exist, remain compatible and seek profit; its security does not rely on identifying that hardware or retiring it through emergency changes. Class v6 adopts the 64-register window and retains it across every rotation. Current modelling estimates a 2.5x to 3.0x energy-efficiency advantage for the strongest specialised designs assessed against the GPU tier (2.1x to 2.6x on the GPU's own node). The long-program and select-tree proposals were rejected. Economic resistance depends on development cost, deployment economics and productive hardware lifetime; family transitions receive an obsolescence benefit only where a loss of competitiveness is demonstrated; programmable multi-epoch designs are included in the assessment.

+

The labels. The range is modelled, never a lower bound: the synthesis-gated figure with the placement’s expected range folded in (the first placed core came in 64 percent above synthesis, wires and the clock tree); the placed gated core’s row narrows it to one figure each when it lands. The GPU side is measured: an RTX 5080 at its 1,100 MHz core lock, 2.06 microjoules per hash, and an RTX 5090 at its 1,300 MHz lock, 2.33 microjoules per hash, both under class v4, on the project’s own rigs and rented pods, 8 October 2026; the card’s cost of the window is measured too (a rented RTX 5090 and RTX 4090 at stock, 8 October 2026: within 5 percent per load with the liveness chain, no register spill). The chip side is synthesised and claimed: the clock-gated sequencer core with the 64-register window on ASAP7, scaled to N3 on the foundry’s headline factors (k about 0.37 at N3 and 0.51 node for node for the base core, the gated window adding about 0.13 of k against an adversary with a flop register file; the window’s liveness measured at 61 of 64 values necessary, its cost to the card measured under 5 percent); the window’s k is synthesis-derived and not a lower bound, and the multi-family adversary lane’s first core (its state in a macro) reads the window’s defence as close to nothing, a disagreement between two models that the placed rows settle. The chip’s memory is modelled: the GDDR7 board of the chip model. The placed gated figure is expected near 2.6x to 3.1x a node ahead and 2.3x to 2.7x node for node (approximate) and is served when its row lands.

Three statements, kept separate. The baseline is the hash as it stands under the scoring rule; rotation is an optional improvement to that baseline, not the mechanism the claim rests on.

- - + +
StatementWhat it saysLabel and date
Energy resistanceThe bracket above: about 2.3x to 3.3x for the strongest specialised design a node ahead of the GPU tier, 2.0x to 2.9x on the GPU’s own node, provisional until the placed gated core row lands; two nodes ahead follows from that row. The honest tier moves to the next node with every GPU generation; a chip must tape out again. Whole machine per tier (synthesis with the SRAM band and node factors, claimed; placed rows to follow): the complete GDDR7 machine about 1.8x the RTX 5090 at its lock per joule node for node and 2.1x a node ahead; 1.6x and 1.9x the RTX 5080; 2.8x and 3.3x the Ada, Ampere and RX 9070 XT cohort; about 1.5x the Apple tier, reported, never headlined.modelled on measured cards, 8 October 2026, approximate and provisional; the node column is claimed scaling; the per-tier line claimed, 8 October 2026
Economic resistanceWhether a chip gets built depends on development cost, deployment economics and productive hardware lifetime. The first cut of the profitability surface: the price at which a project pays scales as the project cost over its share of the chain times its discounted life, and moves by under 5 percent with the per-joule edge; a fixed-lane chip under rotation needs 4x the price a programmable one needs. No threshold is the headline: the five-year coexistence model (Deliverable 4; its first run is docs/analysis/class-v6/coexistence-model.md, every row modelled) replaces any capex wall: the GDDR7 board passes all six of its success conditions at a one to three year life; an N2 SRAM die fails five once it exists with development sunk, and the only condition holding it is that nobody pays to build it; the larger half of a chip's edge is capital cost per accepted hash, not joules.modelled, the coexistence model's first run, 8 October 2026
Energy resistanceThe range above: 2.5x to 3.0x for the strongest specialised design a node ahead of the GPU tier, 2.1x to 2.6x on the GPU’s own node, the placed row to narrow it; two nodes ahead follows from that row. The honest tier moves to the next node with every GPU generation; a chip must tape out again. Whole machine per tier (synthesis with the SRAM band and node factors, claimed; placed rows to follow): the complete GDDR7 machine about 1.8x the RTX 5090 at its lock per joule node for node and 2.1x a node ahead; 1.6x and 1.9x the RTX 5080; 2.8x and 3.3x the Ada, Ampere and RX 9070 XT cohort; about 1.5x the Apple tier, reported, never headlined.modelled on the measured RTX 5090 lock row, 8 October 2026, the chip side synthesised and claimed; the node column is claimed scaling; the per-tier line claimed, 8 October 2026
Economic resistanceWhether a chip gets built depends on development cost, deployment economics and productive hardware lifetime. The first cut of the profitability surface: the price at which a project pays scales as the project cost over its share of the chain times its discounted life, and moves by under 5 percent with the per-joule edge; a fixed-lane chip under rotation needs 4x the price a programmable one needs. No threshold is the headline: the five-year coexistence model (Deliverable 4; its first run is docs/analysis/class-v6/coexistence-model.md, every row modelled) replaces any capex wall: the GDDR7 board passes all six of its success conditions at a one to three year life; an N2 SRAM die fails five once it exists with development sunk, and the only condition holding it is that nobody pays to build it; the larger half of a chip's edge is capital cost per accepted hash, not joules. The result, as the model words it: The test is whether a specialised supplier can earn a normal return while ordinary GPUs stay close enough in total cost, obtainable and useful outside mining that new entrants still compete. On today's modelled rows that holds for the chip anyone can build, a stored-dataset board on commodity DRAM, at a productive life of one to three years: its cost per accepted unit of work sits within the range of the best GPU owner and entrant, and a fleet of it holds a minority of the network with GPU entrants still setting the price. It fails for the SRAM-store die once that die exists with its development paid: no electricity price keeps a GPU owner level with it, and a modest fleet takes most of the network. What holds the die is the investment decision, not the hash; every chip figure here is modelled, not measured, and the chip's hardware cost per unit of work is approximate within 2x.modelled, the coexistence model's first run, 8 October 2026
Response capabilityRotation is an optional improvement, not the mechanism. A passed rotation boundary proves the rotation works, not that hardware dies. The schedule: a new program every hour, a parameter era every week, a family epoch every 180 days, an emergency vote when miners call one. Rotation costs a chip versatility, not life: the family bank is firmware plus about 43 percent of core cells, and no transition carries an obsolescence credit (modelled).measured per boundary, 8 October 2026; the family-bank cost modelled, 8 October 2026

What a miner sees from this. Class v4 costs a 5090 145 W more unlocked, 88 W more at a 1,400 MHz core lock and 82 W at the best operating points (class v4 at 1,200 MHz, class v3 at 1,300; the knee is 1,300 MHz on both), for 0.2 percent more rate (measured, 7 October 2026; the 80 W read on 6 October was at the app's tuned cap); an M5 Max 16 W more for 1.5 percent, an RX 9070 XT and an RTX 4070 nothing (measured, 6 October 2026). The ladder that sets how much work rides in the shadow starts at rung 0 at genesis and climbs by miner signal; its third rung is inadmissible today because a server core verifies it in 10.85 ms, over the gate (measured, 7 October 2026). The devnet runs class v4 from its first block. Classes rotate on findings and at least yearly; a class change is a release activated by block height. Class v5 is built to cross by height; its dataset keyed by the chain's own state is under evaluation (Deliverable 3) and is not counted as a defence until justified. Class v6 is the design in progress (opened 8 October 2026), with four layers as its spine: per-era draws of the parameters a release now fixes, a dataset whose size tracks the chain state (under evaluation, Deliverable 3), scheduled family epochs by height, and the acceptance floor generalised to every era’s draw. The next test of the model is an internal adversarial pass, not an independent review: three lanes that have never worked on the hash code attack the mixer, the chained cache and the acceptance rule with only what an outsider has (the public kit, the frozen object, the spec, the harnesses) and publish the break or the bound they reach. No outside review has run yet.

@@ -490,7 +490,7 @@ body.all .pager{display:none} Light verification256 MB cache on a CPU, milliseconds256 MB cache on a CPU (512 MB from year 4), one warp under 10 ms, the gate. Measured 2.1 ms on one Apple M5 Max core for class v3 (3.4x class v2's 0.61 ms); a 2019-class core not yet Changes over timeNone. A fixed design, unchanged for seven yearsA new program every hour, its memory pattern with it; era draws and reserved families on a schedule fixed at genesis. Nobody touches it Seed grindingNot applicable, the program comes from the hash inputClosed by a verifiable delay between seed and program - Useful workNone. Hashing onlyNVIDIA cards prove: from 8 GB on the patched server, 12 GB and up beside the miner, 24 GB on the stock server (measured on eleven rented cards, RTX 3060 to RTX 5090, 6 October 2026). Selling proofs to other chains is Designed, not built. AMD and Apple cards mine and do not prove; a prover for them lands when a zkVM ships one + Useful workNone. Hashing onlyNVIDIA cards prove: from 8 GB alone on the patched server, 16 GB and up beside the miner at the 5.5 GiB dataset floor, 8 and 12 GB cards time-sharing with the miner paused, 24 GB on the stock server (measured on rented cards, 6 and 8 October 2026). Selling proofs to other chains is Designed, not built. AMD and Apple cards mine and do not prove; a prover for them lands when a zkVM ships one Track recordAbout seven years with one shipped chip, Bitmain’s Antminer X5 (September 2023), at 1.46x per joule over a desktop CPU; the one announced beyond it, Bitmain’s Antminer X9, was withdrawn in May 2026 with zero units; RandomX v2 released 25 March 2026, activation pendingZero years. Every number above is measured and logged with the commands that produced it. The specification, reference hash, test vectors and simulators are public now (git.igneum.network/igneum-network/spec). The node, the miner and the wallet follow to the same host as the repository is published
@@ -516,7 +516,7 @@ body.all .pager{display:none}

Igneum blocks are proven with zero-knowledge proofs, and NVIDIA miners produce them. AMD and Apple cards mine; they do not prove today. Proving is a useful GPU workload that is cheaply verifiable by construction. A proof is right or it is not. Wrapped for light clients, a phone checks it in milliseconds; the wrapping cost is a phase two measurement. Measured so far, the certificate half only: the browser verifier on the home page checks a devnet finality certificate, one BLS aggregate signature over 16 keys and 21 header hashes, in 139 to 155 ms cold and 58 to 68 ms warm in a phone-sized tab on a laptop core (5 October 2026). No phone has been measured, and no wrapped block proof exists yet.

How a block gets proven

Blocks carry transactions only and make no claim about state. Every node executes the ordered transactions natively at once, so users see their transaction land in about a second. The execution is then split into shards of a fixed proving cost. Shards are assigned by lot to eight provers for ten seconds, then open to anyone; there is no bond. Provers run them on consumer NVIDIA cards, and the shard proofs are folded by recursive aggregation into one proof for the block. That proof lands on-chain within about a minute at launch (designed). Because the proof computes the state from the ordered sequence, no node accepts a block with a wrong state root. Full nodes also execute every block natively and reject a proof record whose result differs from their own execution, so a forged proof is a light-client problem and never a chain split. Implemented: the native-execution check on every carried proof record, proving v0 on the devnet (specification section 7). The emergency path for a soundness bug in the proof system is a human one: a new proof-system version is written by people and activates only on miner signalling. Invalid transactions are skipped by rule, the way Kaspa skips conflicting spends.

-

Proving is NVIDIA's today: from 8 GB on the patched server, 12 GB and up beside the miner, 24 GB on the stock server. Measured on eleven rented cards, RTX 3060 to RTX 5090, 6 October 2026: the RTX 3060 (12 GB) mines at 23.78 MH/s and proves the v1 shard beside its miner at an 8.9 GB peak in 37.5 s; the RTX 4060 (8 GB) proves it alone at 7.4 GB in 18.4 s; the RTX 4090 (24 GB) proves it on the stock SP1 server in 5.6 s at 17.4 GB. The patched server that fits the smaller cards is not yet in the shipped app. AMD and Apple cards mine. A prover for them lands when a zkVM ships one. These rows are the proving stage only: the full pipeline is judged, inputs, proving, aggregation, verification, payment, memory and the mining income forgone, and a fast shard does not settle it.

+

Proving is NVIDIA’s today; AMD and Apple cards mine and do not prove. At the 5.5 GiB dataset floor the miner holds about 6.1 GiB and a compressed shard proof peaks at about 7.5 GiB, so mining and proving together on one card needs a 16 GB card; 8 GB and 12 GB cards time-share, the app pausing the miner for the proof (measured 8 October 2026 on a rented RTX 3060 12 GB and RTX 4060 8 GB: the miner 6,129 and 6,116 MiB resident, the proof 7,525 and 7,532 MiB at its peak, 13.2 s and 8.2 s a compressed shard with the miner paused; the record is `docs/analysis/class-v6/coexist-rows.md`). The earlier rows of 6 October 2026 on eleven rented cards, RTX 3060 to RTX 5090, were taken beside a 1 GiB dataset miner (1.4 GB resident): the RTX 3060 (12 GB) proved the v1 shard beside that miner at an 8.9 GB peak in 37.5 s; the RTX 4060 (8 GB) proved it alone at 7.4 GB in 18.4 s; the RTX 4090 (24 GB) proved it on the stock SP1 server in 5.6 s at 17.4 GB. What changed is the miner’s dataset, not the prover. The patched server that fits the smaller cards is not yet in the shipped app. AMD and Apple cards mine. A prover for them lands when a zkVM ships one. These rows are the proving stage only: the full pipeline is judged, inputs, proving, aggregation, verification, payment, memory and the mining income forgone, and a fast shard does not settle it.

The proving budget

Gas prices execution. Proving cost is a different number, so Igneum meters it separately: every transaction pays in both dimensions, and each block has a proving-cost budget set in consensus from measured prover throughput. A transaction that is cheap to run and expensive to prove pays for what it costs the provers. Measured on 5 October 2026 (an RTX 5090 under SP1 6.8.1's GPU prover, the shard size the chain adopts from its fee switch, 30,000 proving gas, about 4.7 million prover cycles): one full shard proves in 4.3 seconds and needs 20.4 GB of GPU memory with the card to itself, so a 24 GB card proves full shards and a 12 GB or 16 GB card does not on this prover build, whose floor is 13.9 GB for even an empty shard; mining and proving on one card needs 32 GB today (the prototype-size shard beside the miner peaked at 30.1 GB) and 24 GB once the adopted shard size is live (22.2 GB beside the miner, 13.2 seconds a shard, measured on the 32 GB card; a 24 GB card has not run it yet). The old 12 GB gate on the roadmap was withdrawn on 5 October until a prover build with a smaller floor was measured; on 6 October a patched server proved the same shard at 7.4 to 8.0 GB alone on eleven rented cards from the RTX 3060 to the RTX 5090 (the real-card table), so the gate returns as measured and the patched server is not yet in the shipped app. The first proofs exist: on 4 October 2026 an RTX 5090 proved a small two-transaction block in 1.4 seconds (2.7 seconds compressed), verified in 0.22 and 0.038 seconds, and a laptop CPU proved a three-shard block end to end in 19 minutes. Later that day the same card proved a full shard at the provisional size, 6.75 million prover gas, which executed in 60.8 million cycles: core proof 8.3 seconds, compressed proof 10.9 seconds, verified in 0.040 seconds; a four-shard block took 44.5 seconds of GPU stages end to end. Shards are assigned and proven on the devnet from block zero. The gate asks for a mid-range card, and an RTX 5090 is not one, so the gate stands open. Once the gate is measured, the budget rises by schedule as hardware improves. The proof system is hash-based, which is what runs on consumer cards, and sits behind a versioned interface. SP1 is the one backend. A replacement is adopted only where justified, by a miner-signalled release, never as an interchangeable second backend, and the chain runs for ever on the current one if none is adopted.

Proving for everyone else

@@ -655,7 +655,7 @@ body.all .pager{display:none}

The size of that third stream today, in numbers: all of Ethereum L1's proving is about USD 36 a day at the September 2026 tracker cost (USD 0.005 a block, 7,200 blocks a day; the tracker figure is a secondary source), against about USD 13,700 a day of Igneum's year-1 emission at USD 0.005 per IGN (31.688 IGN a block, 86,400 blocks a day; the price is an input, not a forecast). So external proving is a small second income at launch and the lottery pays the bills; for proving to become the main income the paid demand would have to grow about 1,000x in dollars (the chip and economy analysis of 6 October 2026, section 3.11; ledger E19).

The honest bear-market case rests on cost. A miner's card is already running and the power is often domestic, so Igneum miners' electricity cost in the proving market is close to power. The price they must charge is another matter: the price a prover must charge is the subsidy it forgoes while it proves, which falls as one over network hash, so the edge over data-centre provers appears only once the network's hash is large (near 100 GH/s for a card proving beside its miner) and is nothing more. Which of the two in-chain streams pays more per GPU-second depends on the size of the fleet: measured on 4 October 2026, three machines at 275 million hashes a second, a second of hashing paid about 4.9x a second of proving the pool share; at 10,000 cards the same arithmetic favours proving by about 930x. That is arithmetic on measured devnet rates, approximate, not a market measurement.

Hardware

-

The dataset starts at 2 GB and grows (the proposed schedule, fixed at genesis: 2 GB, doubling at years 4, 12 and 28, the average of half a gigabyte a year), so a 4 GB card mines for about four years and an 8 GB card for about twelve, approximate. NVIDIA cards prove: from 8 GB on the patched server, 12 GB and up beside the miner, 24 GB on the stock server (measured on eleven rented cards, RTX 3060 to RTX 5090, 6 October 2026). NVIDIA and AMD cards both mine, because the mining program is generated for the architecture both share; only NVIDIA cards prove today. Apple's chips are GPUs with unified memory, so Macs mine too, at about a fifth of a flagship card: Measured, 26.7 against 123 million hashes a second, an Apple M5 Max beside an RTX 5090, mining side by side, 4 October 2026. A Mac is a poor miner per dollar. There is no CPU mining lane, on purpose, because CPU mining is what botnets farm. Nodes, wallets and exchanges need no GPU at all.

+

The dataset starts at 2 GB and grows (the proposed schedule, fixed at genesis: 2 GB, doubling at years 4, 12 and 28, the average of half a gigabyte a year), so a 4 GB card mines for about four years and an 8 GB card for about twelve, approximate. NVIDIA cards prove: from 8 GB alone on the patched server, 16 GB and up beside the miner at the 5.5 GiB dataset floor, 8 and 12 GB cards time-sharing with the miner paused, 24 GB on the stock server (measured on rented cards, 6 and 8 October 2026). NVIDIA and AMD cards both mine, because the mining program is generated for the architecture both share; only NVIDIA cards prove today. Apple's chips are GPUs with unified memory, so Macs mine too, at about a fifth of a flagship card: Measured, 26.7 against 123 million hashes a second, an Apple M5 Max beside an RTX 5090, mining side by side, 4 October 2026. A Mac is a poor miner per dollar. There is no CPU mining lane, on purpose, because CPU mining is what botnets farm. Nodes, wallets and exchanges need no GPU at all.

What a miner's hour looks like

The card hashes the lottery continuously. On an NVIDIA card, when the client sees a shard it can win (or, once the job market is built, an external job), it switches the card to proving for a few seconds, posts the proof, and goes back to hashing. The client does the switching and the miner sees one balance.

The protocol carries no fee: no dev fund, no cut to any team. Ember, the miner software, takes an optional 1% dev fee, the way other GPU miners do. One block template in 100 is requested with the dev address instead of yours, by a counter, not a random draw, so it is exactly 1 in 100 and anyone can check it from the source or from the chain. One flag turns it off (--dev-fee 0, a switch in the app, a line in the HiveOS config). The miner prints the fee and the address when it starts. Any other client is welcome.

@@ -831,7 +831,7 @@ body.all .pager{display:none}

Here are the limits, stated before anyone else states them.

@@ -436,7 +436,7 @@ pre b{color:var(--molten-text);font-weight:500}
Does this website use my GPU to mine?

No. The pictures on the home page are drawn, not mined. Mining happens only in the app you install, and only when you press Start.

-
Does my card mine and prove?

Every card mines. Proving the full shard needs a 24 GB NVIDIA card; a 32 GB card does both at once. Apple silicon proves on the CPU, slowly. The Prove page of the app says in one sentence what your card can do.

+
Does my card mine and prove?

Every card mines. NVIDIA proves; AMD and Apple mine. Mining and proving together needs a 16 GB NVIDIA card at the 5.5 GiB dataset floor (the miner about 6.1 GiB, a compressed proof about 7.5 GiB at its peak); an 8 GB or 12 GB card time-shares, the app pausing the miner for the proof; on the stock server the full shard needs 24 GB and a 32 GB card does both at once (measured, 6 and 8 October 2026). Apple silicon proves on the CPU, slowly. The Prove page of the app says in one sentence what your card can do.

Is the devnet paying real money?

No. Devnet coins have no value and the chain may reset. The app’s pounds row reads 0.00 on devnet and says why. The Igneum 2.0 devnet is the network today. Mainnet has not started.

Is there a fee?

Not in the protocol: no dev fund, no fee to any team. The app takes an optional 1% software fee, the norm for GPU miners, and one flag turns it off. The fee, in full view.

Can I run it on a rig or in a pool?

The Linux and HiveOS tarball is above, with the flight sheet. Today every machine mines solo on its own keys, and a card runs several.

diff --git a/site/randomx.html b/site/randomx.html index 366e99e61..531902d66 100644 --- a/site/randomx.html +++ b/site/randomx.html @@ -254,7 +254,7 @@ Light verification256 MB cache on a CPU, milliseconds256 MB cache on a CPU (512 MB from year 4), one warp under 10 ms, the gate. Measured 2.1 ms on one Apple M5 Max core for class v3 (3.4x class v2's 0.61 ms); a 2019-class core not yet Changes over timeNone. A fixed design, unchanged for seven yearsA new program every hour, its memory pattern with it; era draws and reserved families on a schedule fixed at genesis. Nobody touches it Seed grindingNot applicable, the program comes from the hash inputClosed by a verifiable delay between seed and program - Useful workNone. Hashing onlyNVIDIA cards prove: from 8 GB on the patched server, 12 GB and up beside the miner, 24 GB on the stock server (measured on eleven rented cards, RTX 3060 to RTX 5090, 6 October 2026). Selling proofs to other chains is Designed, not built. AMD and Apple cards mine and do not prove; a prover for them lands when a zkVM ships one + Useful workNone. Hashing onlyNVIDIA cards prove: from 8 GB alone on the patched server, 16 GB and up beside the miner at the 5.5 GiB dataset floor, 8 and 12 GB cards time-sharing with the miner paused, 24 GB on the stock server (measured on rented cards, 6 and 8 October 2026). Selling proofs to other chains is Designed, not built. AMD and Apple cards mine and do not prove; a prover for them lands when a zkVM ships one Track recordAbout seven years with one shipped chip, Bitmain’s Antminer X5 (September 2023), at 1.46x per joule over a desktop CPU; the one announced beyond it, Bitmain’s Antminer X9, was withdrawn in May 2026 with zero units; RandomX v2 released 25 March 2026, activation pendingZero years. Every number above is measured and logged with the commands that produced it. The specification, reference hash, test vectors and simulators are public now (git.igneum.network/igneum-network/spec). The node, the miner and the wallet follow to the same host as the repository is published