← All guides

Underclocking for efficiency: power limits, J/TH and the break-even rule

Costs & math · 9 min read

Educational content only, not financial advice. Worked examples below use other people’s published figures and clearly labelled illustrations to teach the arithmetic; they are not predictions of what any specific machine will earn. Run the same arithmetic with your own meter readings and your own electricity rate before changing any setting.

Stock mining settings are chosen by manufacturers to look good on a specification sheet, which usually means maximum hashrate. The miner paying the electricity bill wants a different thing: maximum hashrate per watt. Underclocking, running chips at lower frequency and voltage than stock, and its GPU cousin, the power limit, exist because the trade between the two is not one-for-one. Power falls faster than hashrate as settings drop, so a slower machine can be a cheaper machine to run per unit of work. This guide verifies the arithmetic on published examples, gives a tuning procedure that works on both ASICs and GPUs with stability windows and a rollback rule, and then states the part most tuning guides omit: the conditions under which underclocking makes efficiency better and profit worse.

The one metric: efficiency, defined two ways

For ASICs the standard metric is joules per terahash (J/TH): watts divided by terahashes per second. Lower is better. For GPUs the same idea is usually written as hashrate divided by power, megahashes per watt, higher is better, and tuning guides state the rule identically: measure hashrate divided by power consumption, because raising power sometimes raises hashrate by less than the extra electricity justifies. An illustrative example from a tuning guide (its numbers are labelled examples, not a specific card’s promise) makes the shape clear:

ConfigurationHashrate (units)PowerEfficiency (hashrate ÷ power)
Stock100250 W0.40
Reduced power94190 W0.49
Optimised97200 W0.49

Hashrate fell 6% while power fell 24%, so efficiency rose by about a fifth. Every claim in underclocking is some version of that table, which is why every figure in this guide is recomputed rather than quoted on trust.

Why watts fall faster than hashes

Chip power consumption rises non-linearly with clock speed, because higher frequencies demand higher voltage, and the voltage relationship is roughly quadratic: reducing voltage by about 10% can reduce power consumption by roughly 19%, while hashrate falls by closer to the frequency reduction alone. That asymmetry is the entire physical basis of underclocking. It also explains its limit: at some point the curve flattens, further reductions cost hashrate one-for-one or cause instability, and the efficiency gains stop. The sweet spot is a property of each individual machine’s silicon, which is why per-chip tuning exists, and why copied settings from another machine are a starting guess, never a conclusion.

Worked ASIC example, recomputed

A widely cited tuning example uses an Antminer S19j Pro rated at 104 TH/s and 3,068 W. Check the stock efficiency: 3,068 ÷ 104 = 29.5 J/TH, matches the published figure. The same source reports the machine underclocked to approximately 75 TH/s at around 1,800 W: 1,800 ÷ 75 = 24 J/TH, also matches. In percentage terms, that is 28% of hashrate surrendered for 41% of power saved. A second published example, an S19 Pro taken from 110 TH/s and 3,250 W (29.5 J/TH) down to 80 TH/s and 2,200 W, recomputes to 27.5 J/TH, a real but smaller gain, which is the honest range to expect: different machines, different silicon, different sweet spots. Firmware is the enabler on ASICs: stock firmware on many models offers little or no power targeting, while aftermarket firmware such as BraiinsOS, VNish and LuxOS adds per-chip tuning, power targets and dynamic frequency adjustment. That capability is not free of trade-offs, aftermarket firmware typically carries a developer fee, and tuning beyond manufacturer settings can affect warranty coverage; overclocking is documented to void most warranties, and underclocking through unofficial firmware should be assumed to carry the same warranty question until the manufacturer’s terms say otherwise.

GPUs: power limit first, clocks second

On a GPU the same physics is reached through different controls, and the order matters:

Power limit is the primary lever. Limiting the card’s total consumption is what increases mining efficiency; one documented guide workflow sets the limit to 65%, in MSI Afterburner as a slider, or in the miner’s own arguments (for example --pl 65 in GMiner, noting that some miners such as NBMiner expect watts rather than a percentage).

Core clock usually goes down, not up. Most mining algorithms are memory-bound rather than core-bound, so lowering the core clock cuts power and heat with little hashrate cost, while memory clock is raised cautiously where the algorithm rewards it, one documented range for the Octopus algorithm is +800 to 1000 MHz on RTX 3000 to 4000 series cards and +500 to 750 MHz on RTX 2000 series, dialled in gradually.

Heat is the constraint that binds. The same guide workflow aims to keep cards at or below about 65°C, with fan speed around 40 to 60% as the starting point. A power limit that only works with fans at maximum has traded an electricity saving for a fan-replacement schedule.

Tuning procedure: small steps, long windows, one rollback

1. Baseline for 24 hours at stock. Record hashrate, wall power (measured, not read from software where a meter is available), temperatures, fan speeds and rejected-share rate. Compute the baseline efficiency (J/TH or hashrate ÷ watts). Every later claim is measured against this line.

2. Change one variable, a little. Reduce the power target or power limit by about 5%, or trim voltage in the 3 to 5% steps tuning guidance recommends as a starting increment. Never change power, core clock and memory clock in the same step, if stability changes, the cause must be identifiable.

3. Hold each step for a real stability window. Published tuning guidance tests each adjustment for 48 to 72 hours and watches rejected shares carefully, because rising errors silently negate the savings. A setting that survives an hour has proven almost nothing at 24/7 duty.

4. Recompute efficiency at each step from measured wall power and average hashrate over the window, not from the best hour, and log it next to the setting that produced it.

5. Stop at the knee. When a further 5% power cut costs more than about 5% hashrate, or rejected shares climb, the sweet spot is behind you: restore the previous logged setting. That restore is the rollback, and it should be a known value, not a memory.

6. Re-check seasonally. A tune that is stable in a cool room can throttle or error in summer heat; conservative warm-weather profiles and aggressive cool-weather profiles are standard practice, not indecision.

What success looks like: after the final setting has run a full 48 to 72-hour window, measured efficiency is better than the step-1 baseline, rejected shares are no higher than baseline, temperatures are at or below their target with fan headroom to spare, and the daily electricity cost (watts ÷ 1,000 × 24 × your rate) has fallen by more, in currency, than the value of the hashrate given up. That last clause is the subject of the next section.

When underclocking loses money

Efficiency is a ratio; profit is a subtraction. Underclocking always improves the ratio within its useful range, but it reduces total hashrate, and whether that trade pays depends on the margin per hash at the current electricity price. The decision framework:

SituationVerdictReasoning
Machine is unprofitable at stock, or barely profitable, at the actual electricity rateUnderclock, the strongest casePower saved is worth more than the hashes surrendered; underclocking can keep older hardware viable that stock settings would retire
Machine is comfortably profitable per hash at a low rateUnderclock lightly or not at allEvery surrendered hash was earning more than its power cost; deep cuts destroy margin to save cheap watts
Heat, noise or a circuit limit is the binding constraintUnderclock to the constraintThe alternative is the machine off or throttling; a lower steady hashrate beats an unstable higher one, and less heat extends hardware life
Warranty is live and valuableStock settings, or manufacturer power modes onlyEfficiency gains are small against a voided warranty on a hashboard

The published S19j Pro example is explicitly framed the same way by its source: at higher electricity rates the trade-off is strongly favourable; by implication, at very cheap rates it weakens. Note also what underclocking is not: it does not rescue a machine whose best possible efficiency still loses money at the local rate. It moves the break-even electricity price upward, sometimes enough, sometimes not.

FAQ

What is J/TH? Joules per terahash: watts divided by hashrate in TH/s. The stock S19j Pro figure of 29.5 J/TH is simply 3,068 W ÷ 104 TH/s, recomputed above.

How much power can underclocking save? Published worked examples save 32 to 41% of power for 27 to 28% of hashrate. Individual machines vary; the procedure above exists because the sweet spot cannot be copied, only found.

Does a GPU power limit hurt hashrate much? On memory-bound algorithms, modestly, the illustrative tuning table shows a 24% power cut costing 6% of hashrate. Core-heavy algorithms punish core cuts more, which is why one variable changes at a time.

Is aftermarket firmware required? For ASIC power targets and per-chip tuning, usually, stock firmware is comparatively limited. Weigh the developer fee and the warranty position before treating it as a free upgrade.

Next: Mining rig maintenance schedule →