Definition
Overclocking is the practice of running mining hardware above its factory clock frequency to extract more hashrate, almost always paired with a voltage increase to keep the ASIC chips stable at the higher speed. The reward is more terahashes per second; the cost is higher power draw, more heat, louder fans, and a shorter hardware lifespan.
Also known as: OC.
In an ASIC miner, hashrate scales with chip frequency. Each Bitmain chip contains a PLL (phase-locked loop) that multiplies a 25 MHz reference clock up to the operating frequency, and overclocking means writing a higher divider value into that PLL register. On the BM1387 used in the S9 the frequency is set by four divider fields, so firmware does not write “MHz” directly; it picks a PLL-table entry resolving to the nearest achievable frequency. Stock firmware applies one uniform frequency to every chip on a hashboard, but custom firmware can address chips individually and write per-chip PLL values.
Voltage, however, is not per chip. On modern Antminer boards a group of roughly ten chips shares a single DC-DC voltage domain, so raising voltage happens per-domain, never per individual ASIC. Confusing per-chip frequency with per-domain voltage is one of the most common mistakes in mining writeups.
How overclocking actually works in firmware
Naively cranking the frequency knob produces hardware errors: a chip clocked beyond its silicon’s stable point starts returning bad nonces, and the error rate climbs. Good firmware handles this with an autotuner rather than guesswork. A characterization pass searches for the highest frequency each chip can hold while keeping its per-chip error rate under a threshold — a 2% error rate per measurement window is a typical cutoff. A second pass walks voltage downward to find the minimum that still supports that frequency, then adds a small safety margin (around 20 mV) so normal drift does not tip a chip into errors.
This multi-pass approach is why a real tune takes time. Firmware families converge at different speeds — some gradient-descent tuners take well over an hour on a high-chip-count machine, while binary-search and adaptive tuners are faster — but all do the same work: characterize each chip, find a stable voltage, then soak at temperature to confirm the operating point holds once the board is hot. A tune that skips the soak looks excellent cold and degrades as the boards warm up.
It is also why power-targeting firmware does not ask for a raw MHz value at all. You select a wattage or hashrate target and the autotuner derives the frequency and voltage at runtime, because silicon quality varies from unit to unit and even chip to chip. The firmware stores only watt-to-hashrate pairs; the exact frequency and voltage are computed live for your machine.
The efficiency tradeoff, with real numbers
Overclocking buys hashrate but sells efficiency. Watts rise faster than hashrate, so efficiency (J/TH) gets worse the harder you push. A 126 TH-class S19 (BM1398) shows the spectrum: a conservative profile runs roughly 1630 W for 67 TH at about 24.3 J/TH, while an aggressive overclock reaches roughly 4700 W for 130 TH at about 36.2 J/TH — nearly triple the power for under 2x the hashrate, a deliberately uneconomic trade you make only when raw terahashes matter more than your power bill.
Cooling, limits, and the real tradeoffs
Overclocking is fundamentally a thermal problem. Pushing frequency and voltage raises die temperature, and silicon that runs hot ages faster and errors more. Firmware enforces guardrails: a typical autotuner begins derating frequency once die or board temperature crosses a threshold near 60°C and triggers an emergency shutdown around 75°C to protect the boards. If cooling cannot keep up, the firmware quietly claws back the frequency you gained, so a serious overclock demands serious airflow — high-static-pressure fans, clean intakes, and for the most aggressive profiles, immersion cooling or hydro setups that move more heat than air ever can.
An example from the small end of the scale: a BM1387-based S9 ships at 600 MHz, runs comfortably around 650 MHz, and is considered overclocked at 700 MHz. The jump sounds modest, but the extra heat and power are very real, and the chips that survive it have both good silicon and good cooling. The same principle scales up to current-generation hardware — only the numbers change.
- Hashrate gain: more terahashes, more potential block rewards.
- Power cost: watts rise faster than hashrate, so J/TH efficiency gets worse.
- Heat and noise: both climb sharply and demand better cooling.
- Lifespan: sustained high voltage and temperature accelerate chip aging.
Overclocking is the mirror image of undervolting and underclocking, which trade hashrate for better efficiency; which direction you tune depends on whether your bottleneck is hardware or electricity. Either way the capability lives in the firmware — stock firmware exposes little of it, while custom firmware unlocks per-chip frequency control and full autotuning. Open-source rigs like the Bitaxe make this approachable: the AxeOS web interface lets you set frequency and core voltage directly and watch the error rate respond in real time, making a single-ASIC board an excellent place to learn how overclocking behaves before you risk a full-size machine.
If you are weighing which direction to tune your fleet, or which hardware tolerates an aggressive profile best, our firmware comparison and the catalog of ASIC miners in the D-Central shop are the place to start. And if a board starts throwing errors after a push, the ASIC troubleshooting guide will help you walk it back.
In Simple Terms
Running mining hardware faster than default to increase hashrate, at the cost of more power and heat.
Overclocking is the practice of running mining hardware above its factory clock frequency to extract more hashrate, almost always paired with a voltage increase to keep the ASIC chips stable at the higher speed. The reward is more terahashes per second; the cost is higher power draw, more heat, louder fans, and a shorter hardware lifespan.
Also known as: OC.
In an ASIC miner, hashrate scales with chip frequency. Each Bitmain chip contains a PLL (phase-locked loop) that multiplies a 25 MHz reference clock up to the operating frequency, and overclocking means writing a higher divider value into that PLL register. On the BM1387 used in the S9 the frequency is set by four divider fields, so firmware does not write “MHz” directly; it picks a PLL-table entry resolving to the nearest achievable frequency. Stock firmware applies one uniform frequency to every chip on a hashboard, but custom firmware can address chips individually and write per-chip PLL values.
Voltage, however, is not per chip. On modern Antminer boards a group of roughly ten chips shares a single DC-DC voltage domain, so raising voltage happens per-domain, never per individual ASIC. Confusing per-chip frequency with per-domain voltage is one of the most common mistakes in mining writeups.
How overclocking actually works in firmware
Naively cranking the frequency knob produces hardware errors: a chip clocked beyond its silicon’s stable point starts returning bad nonces, and the error rate climbs. Good firmware handles this with an autotuner rather than guesswork. A characterization pass searches for the highest frequency each chip can hold while keeping its per-chip error rate under a threshold — a 2% error rate per measurement window is a typical cutoff. A second pass walks voltage downward to find the minimum that still supports that frequency, then adds a small safety margin (around 20 mV) so normal drift does not tip a chip into errors.
This multi-pass approach is why a real tune takes time. Firmware families converge at different speeds — some gradient-descent tuners take well over an hour on a high-chip-count machine, while binary-search and adaptive tuners are faster — but all do the same work: characterize each chip, find a stable voltage, then soak at temperature to confirm the operating point holds once the board is hot. A tune that skips the soak looks excellent cold and degrades as the boards warm up.
It is also why power-targeting firmware does not ask for a raw MHz value at all. You select a wattage or hashrate target and the autotuner derives the frequency and voltage at runtime, because silicon quality varies from unit to unit and even chip to chip. The firmware stores only watt-to-hashrate pairs; the exact frequency and voltage are computed live for your machine.
The efficiency tradeoff, with real numbers
Overclocking buys hashrate but sells efficiency. Watts rise faster than hashrate, so efficiency (J/TH) gets worse the harder you push. A 126 TH-class S19 (BM1398) shows the spectrum: a conservative profile runs roughly 1630 W for 67 TH at about 24.3 J/TH, while an aggressive overclock reaches roughly 4700 W for 130 TH at about 36.2 J/TH — nearly triple the power for under 2x the hashrate, a deliberately uneconomic trade you make only when raw terahashes matter more than your power bill.
Cooling, limits, and the real tradeoffs
Overclocking is fundamentally a thermal problem. Pushing frequency and voltage raises die temperature, and silicon that runs hot ages faster and errors more. Firmware enforces guardrails: a typical autotuner begins derating frequency once die or board temperature crosses a threshold near 60°C and triggers an emergency shutdown around 75°C to protect the boards. If cooling cannot keep up, the firmware quietly claws back the frequency you gained, so a serious overclock demands serious airflow — high-static-pressure fans, clean intakes, and for the most aggressive profiles, immersion cooling or hydro setups that move more heat than air ever can.
An example from the small end of the scale: a BM1387-based S9 ships at 600 MHz, runs comfortably around 650 MHz, and is considered overclocked at 700 MHz. The jump sounds modest, but the extra heat and power are very real, and the chips that survive it have both good silicon and good cooling. The same principle scales up to current-generation hardware — only the numbers change.
- Hashrate gain: more terahashes, more potential block rewards.
- Power cost: watts rise faster than hashrate, so J/TH efficiency gets worse.
- Heat and noise: both climb sharply and demand better cooling.
- Lifespan: sustained high voltage and temperature accelerate chip aging.
Overclocking is the mirror image of undervolting and underclocking, which trade hashrate for better efficiency; which direction you tune depends on whether your bottleneck is hardware or electricity. Either way the capability lives in the firmware — stock firmware exposes little of it, while custom firmware unlocks per-chip frequency control and full autotuning. Open-source rigs like the Bitaxe make this approachable: the AxeOS web interface lets you set frequency and core voltage directly and watch the error rate respond in real time, making a single-ASIC board an excellent place to learn how overclocking behaves before you risk a full-size machine.
If you are weighing which direction to tune your fleet, or which hardware tolerates an aggressive profile best, our firmware comparison and the catalog of ASIC miners in the D-Central shop are the place to start. And if a board starts throwing errors after a push, the ASIC troubleshooting guide will help you walk it back.