Guide
Benchmarking and mining without a pool
A real stratum server inside the miner, so a measured share is an accepted one.
Two ways to make the miner work without pointing it at anyone, answering two different questions. Neither needs a wallet, an internet connection or an account.
Does this machine mine this algorithm?
--bench mines a job made up locally. No sockets at all. It exercises the kernels and
the CPU path and nothing else, so it is the quickest way to find out whether a card and a driver
can run an algorithm before you commit a wallet to it.
./bzminer --bench -a ergo
./bzminer --bench -a kawpow --nvidia # one vendor only
./bzminer --bench -a randomx --cpu_threads 8Every algorithm accepts it. start_sha256d in the archive is this for the SDK example, which has no pool worth pointing at.
How fast is it, really?
--benchmark is the measurement tool. bzminer starts a real stratum server inside
itself, on a local port, then connects to it over TCP like any other pool. The whole client path
runs: subscribe, authorise, jobs, share submit, the pool’s verdict, the accepted / rejected /
stale counters and the pool-hashrate column. It is the honest way to measure a card, because a
share only counts once the server has accepted it.
./bzminer --benchmark -a warthog # default difficulty 1000
./bzminer --benchmark 25000000000000 -a pearl # fixed difficulty, 25T
./bzminer --benchmark 5000 -a xelis --oc-power-limit 200 # measure at a given power capPick the difficulty deliberately
Too low and a fast rig floods the local server with submissions and measures the plumbing instead of the kernel. Too high and you wait a long time for enough shares to mean anything. Aim for a share every few seconds: Pearl on a big GPU wants something like 25T; XELIS wants something in the thousands.
The share count is what to trust
shares × difficulty ÷ seconds is a rate the miner cannot flatter. Compare it
against the displayed hashrate, and if the two disagree for long, one of them is lying. Each
benchmark job is held for ten minutes by default so a memory-hard kernel is not thrown away
mid-batch every time the job refreshes; --benchmark-job-interval 4000 makes jobs go
stale every four seconds if what you want to exercise is how the miner copes with that.
Both modes honour everything else: --nvidia to test one vendor, the --oc-*
flags to measure at a setting, -o log for a scrolling log instead of the dashboard.
Some algorithms register a benchmark screen of their own, which --benchmark opens by
default; -o tui overrides it.
How to compare two settings
- Same session, same machine, nothing else running. Machine state drifts; a number from last week is not a baseline.
- Run the baseline, then the change, then the baseline again. If the two baselines disagree, the machine is not idle enough to measure.
- Long enough for a few hundred shares. A minute of a bursty GPU says very little.
- Change one thing. Two changes at once cannot be told apart afterwards.
- Judge by accepted shares over the window, not the displayed rate.
The overclocking guide says what each algorithm is bound by, which is where to start; this is how to find out whether it helped.
Measuring a narrower CPU path
To find out what an instruction set is actually worth on a rig, tell the miner to pretend it does not have it. Each flag also removes everything built on top of it, so the selected path is the one you meant:
--disable_avx512 # and AMX with it
--disable_avx2 # and VAES and AVX-512, since both need AVX
--disable_vaes
--disable_sse # the portable fallback
--disable_huge_pages # ordinary pages: are huge pages worth anything on THIS rig?Problems
- Lots of rejected or stale shares in
--benchmark. The difficulty is too low for the rig and the local server is being flooded. Raise the difficulty. - Hardly any shares. The difficulty is too high. Lower it until one arrives every few seconds.
- The displayed rate and the share-derived rate disagree. Give it longer; if they still disagree after a few hundred shares, trust the shares and report it.