igneum

History

igneum-labs 9c194ad675 Aggregation cost on the 5090 (5 October, night): the chained aggregation is 2.1 s alone and 9.7 s beside the miner, the batch-log2 curve (2^16 buys 1.6x for a fifth of the hash rate), batch and tree folds estimated, two streams and SP1 knobs closed; the host times the stdin build and names the knobs, --save-shards; the PC 2 job scripts and the readers The statement and the pinned guests are unchanged; every fixture proof verifies as before. The defaults stay (batch-log2 22, SP1 defaults): the one knob that moves a mining card's prover costs a fifth of the hash rate; the plan carries the trade for the project lead and the batch fold for the next pin. Measured: docs/bench-log.md "aggregation cost on the RTX 5090"; the plan line: docs/plans/proving-v1.md "Aggregation cost (5 October, night)". Also: make-package's gate skips the exporter's .node-plan.json side files and takes the run lock for its execute step; the state-reply class (/api/state answering {} once paid_wei passes u64::MAX) found on the way and fixed on the app branch at `e4e9726`. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> (cherry picked from commit ea38ece9eaa5949dd657cbfea5308c4948177ff8)		2026-10-06 07:04:50 +00:00
..
make-package.sh	Aggregation cost on the 5090 (5 October, night): the chained aggregation is 2.1 s alone and 9.7 s beside the miner, the batch-log2 curve (2^16 buys 1.6x for a fifth of the hash rate), batch and tree folds estimated, two streams and SP1 knobs closed; the host times the stdin build and names the knobs, --save-shards; the PC 2 job scripts and the readers	2026-10-06 07:04:50 +00:00
PROVE-BLOCK.bat	Prover package: PROVE-SHARD.bat runs the shard at S_p and the two- and four-shard blocks on the GPU	2026-10-04 10:09:33 +00:00
prove-block.sh	wsl scripts: the re-stamp prunes every target dir, not only one level deep (touching target made cargo skip the rebuild)	2026-10-05 10:51:10 +00:00
PROVE-SHARD.bat	Prover package: PROVE-SHARD.bat runs the shard at S_p and the two- and four-shard blocks on the GPU	2026-10-04 10:09:33 +00:00
prove-shard.sh	Rotation phase 2: packagers read the intake key and the downloads token from files (IGNEUM_INTAKE_KEY_FILE, IGNEUM_DL_TOKEN_FILE, .next by default), no key literal in the tree, app header line with fingerprints, ship-app --dl-both, logs --rotation, tools/repo/fresh-repo.sh with the dry run, docs/plans/rotation-phase-2.md	2026-10-05 07:43:44 +00:00
README.txt	Prover package: PROVE-SHARD.bat runs the shard at S_p and the two- and four-shard blocks on the GPU	2026-10-04 10:09:33 +00:00
SETUP-PROVER.bat	Proving v0: SP1 guest and host for one Igneum block, real-block fixtures, versioned ProofSystem trait, WSL2 package for the RTX 5090 run	2026-10-03 22:30:12 +00:00
setup-prover.ps1	Prover package: strip the UTF-16 nulls from the WSL distro list so an installed Ubuntu is detected	2026-10-04 08:28:39 +00:00
setup-wsl.sh	wsl scripts: the re-stamp prunes every target dir, not only one level deep (touching target made cargo skip the rebuild)	2026-10-05 10:51:10 +00:00

README.txt

Igneum proving on a Windows 11 PC with an NVIDIA card (RTX 5090), through WSL2 Ubuntu 24.04. Devnet v4 shards (4 October 2026).
1. Double-click SETUP-PROVER.bat. First run: it enables WSL2 and installs Ubuntu 24.04 (about 400 MB), then asks you to REBOOT.
2. After the reboot Ubuntu opens once and asks for a username and password (any, e.g. igneum). Close it. Double-click SETUP-PROVER.bat again:
   it installs the CUDA toolkit for WSL (about 3 GB), Rust (about 500 MB), sp1up and the SP1 toolchain (about 300 MB download, 1 GB on disk), then pre-builds the prover.
   Downloads in total: about 4 to 5 GB (approximate). Time: 15 to 40 minutes depending on the connection and the CPU (approximate). The NVIDIA driver is NOT installed inside Ubuntu: Windows provides it. Docker is not needed (SP1 v6 runs a native sp1-gpu-server, which the SDK downloads on first use, about 134 MB).
3. PAUSE MINING (Ctrl+C in the START-MINING window). The prover needs the whole GPU (24 GB of VRAM or more) and most of the RAM.
4. Double-click PROVE-SHARD.bat. This is the phase 2 measurement: it proves one full shard at its proving-gas budget S_p (block-338-shard1, 6.75 M pgas: execute, core proof, compressed proof, each verified), then the two-shard block (block-341-shards2) and the four-shard block near the block budget B_p (block-344-shards4) end to end: a compressed proof per shard and the aggregated block proof by recursion, verified. Every stage prints a STAGE line and a RESULT line with a UTC timestamp. The log is uploaded to the Igneum log intake (run_id shards-<PC name>-<stamp>).
   First run: 10 to 30 minutes of compiling before the first proof (approximate). The proofs: unknown on this card, that is what this run measures (the 4 October morning run proved a 1,488-pgas block in 1.4 s core and 2.7 s compressed; a shard here carries 4,500x the pgas).
5. PROVE-BLOCK.bat still proves the small block-78-increment (mode all: shard 0 in all three stages, then the block end to end), then a shard on the CPU for comparison.
6. Resume mining afterwards. Results also land next to this file in results/*.json. Each line marked RESULT in the log is a number for docs/bench-log.md.
Everything version-specific (CUDA 12.8 package name, SP1 v6.8.1, the server asset name) was read from the SP1 and NVIDIA documentation on 3 October 2026 and is approximate from that date on.