The first support question a new operator asks is always the same. The box says 270 TH/s, the machine’s own web interface says about 270, and the pool says 261. Somebody is lying.
Nobody is. There are four separate gaps between a rated figure and a pool figure, and only two of them are worth doing anything about.
1. Pool-side hash rate is a statistical estimate
The pool cannot see your machine’s hash rate. It sees shares arriving and infers a rate from them, which means the figure wanders — a few per cent either way over an hour, converging over a day. Judge a machine on a 24-hour average, never on a five-minute reading. Half the “underperforming machine” tickets we get resolve themselves by the next morning.
2. Stale and rejected shares
Work submitted after the network has moved on earns nothing. On a well-connected site this is under 0.5%; on a site with a slow or distant pool connection it can reach two or three per cent, which is real money.
This one you can fix. Use a pool with a node near you, keep the latency under about 100 ms, and do not run mining traffic over a saturated office link. It costs nothing and it is the highest-yield thing on this list.
3. Thermal throttling
Every current-generation machine backs off when board temperature rises. On a 40 °C hall day, six per cent off the top is normal for an air-cooled unit; the firmware is protecting an asset you paid for.
Fixable only by fixing the hall — intake temperature, filter condition, airflow path. Loaded filters are the most common cause and the cheapest to address.
4. Tolerance
Rated hash rate carries ±3% at Bitmain and ±5% at MicroBT, and that is a real distribution, not a formality. In a batch of a hundred, some machines land under the rated number and stay there. That is what the tolerance means and it is why the batch test report matters: you are checking that the spread is normal, not that every unit hits the headline.
What tuning actually buys you
Modern firmware exposes a power curve. You can move along it in both directions, and which way depends entirely on your tariff.
Underclocking is the underused half. Backing a machine off by 10% often buys 6–8% in J/TH, because you are moving down the voltage-frequency curve where efficiency improves faster than output falls. Above roughly six cents a kilowatt-hour, that trade is straightforwardly profitable — you produce slightly less and keep more of it.
Overclocking does the reverse and only makes sense on very cheap power, or briefly when the network is unusually favourable. It also runs the boards hotter, which is a warranty conversation waiting to happen.
Third-party firmware is where the real gains and the real risks both live. It voids factory cover, full stop. On a large fleet with in-house repair capability that can be an acceptable trade; on twenty machines with no bench, it is not.
The settings we hand over
Machines shipped with firmware pre-set leave here with the pool configured, the worker naming scheme in place, the power profile matched to the tariff you gave us, and the fan curve left at stock. The batch report records what each unit did on the bench under that profile, so when the pool figure settles a week later you have something to compare it against.
If it settles three per cent low across the whole batch, look at the hall. If one machine in forty sits well below its siblings, that is a board, and that is what the warranty is for.