Skip to content

Operations Mining Acceleration

github-actions[bot] edited this page Nov 21, 2026 · 5 revisions

Mining on the neural development network

The CPU miner evaluates each nonce and computes the scratchpad when the BLAKE3 digest of the seed, nonce and neural score passes the first gate. The CUDA plugin evaluates batches of neural scores. The host checks both digest thresholds and reproduces the winning neural score before signing. Each accepted SOLO submission is a full block candidate. Protocol

Build the node, pool and miner with cargo build --release --locked in their respective repositories. Start all three with --network neural-devnet. Use --no-accelerator --threads 1 for the CPU path or --accelerator-dir DIRECTORY for a CUDA plugin. The miner logs the backend used for each winning solution. A device error retries the affected window on the CPU.

On Windows, python scripts/measure-mining.py --seconds 70 --accelerator-dir DIRECTORY starts a fresh node and SOLO pool per backend. It checks reported heights against the miner log and records valid and invalid shares, executable digests, the CUDA plugin digest and sampled GPU board power under target/mining-measurement-*/. CPU and CUDA start from separate genesis databases and retarget independently. Use an external whole-system meter and fixed challenge windows for energy comparisons. Measurement script

cargo run --release --locked -p hyphen-consensus --example work_cost measures 16 deterministic windows of 1,024 nonces. The CSV records the number of memory evaluations, accepted nonces and time spent independently verifying all winners.

HYN3 short run

Windows 10 build 19045, Core i7-11850H, RTX A4000 Laptop GPU, driver 596.71. Each backend ran with one host mining thread and started from genesis. The miner’s search median covers completed solutions. The board sensor was sampled 138 times per run. The CPU run overlapped other local work. Sampling implementation

Backend Seconds Valid / invalid shares Completed search median GPU board mean GPU board energy
CPU 70.03 21 / 0 99.227 ms 27.17 W 1,893 J
CUDA scores, host scratchpad 70.06 23 / 0 66.829 ms 19.61 W 1,371 J

Both runs crossed the first retarget. Their tips had different cumulative weights and different challenge histories. Board readings exclude host CPU energy and other system loads. Estimating comparative energy per trial requires repeated fixed ranges and a whole-system meter.

Matched nonce windows and live SOLO runs, 2026-11-21

Across 16 matched windows the one-thread Rust path searched 2,504 nonce/s and CUDA scoring with host gates searched 3,048 nonce/s. All 131,072 scores and all 507 winners matched. The CUDA scoring call alone averaged 92,162 scores/s. Setup, timing boundaries and Rust reproduction

Two additional live runs used separate fresh HYN3 databases, a local node and SOLO pool, one mining thread and the same development miner key. Every accepted height had a solution logged by the selected backend. The CUDA log showed no device fallback. Board samples came from nvidia-smi at roughly one-second intervals. Energy below integrates adjacent board readings by timestamp.

Backend Mining seconds Valid / invalid shares Last height GPU samples GPU temperature GPU mean power GPU board energy
CPU 180.26 22 / 0 22 179 55–57 °C 30.30 W 5,464 J
CUDA scoring, host gates 180.08 22 / 0 22 179 50–57 °C 23.29 W 4,193 J

The CPU run also recorded GPU board power from the shared laptop. The forks retargeted separately and local workloads varied. These observations cover three minutes per backend. A sustained thermal qualification and whole-machine energy per matched attempt still require separate measurements. Raw logs, board samples and binary digests are under target/audit-local/live-e213fc7f-e84a-4fe8-b757-90e034a8af71/.

Clone this wiki locally