Clock, Voltage, and Overclocking Risks

On an ASIC, clock (hash-core frequency) and voltage at the chips are the two parameters firmware moves when people say underclock, overclock, or "profile". The official ranges live in the model manual. What follows describes what breaks when you leave them, and how to test a change without losing hardware.

What the two parameters do

Clock sets how often a core attempts a hash per unit of time. Raising it increases the trial rate, and also the heat, the current drawn, and the sensitivity to electrical noise.

Voltage gives the transistors the margin they need to switch cleanly at that clock. Too low for the requested frequency, it produces errors, unstable boards, and a local hashrate that counts failed trials. Too high, it adds heat and current for a frequency gain that may not exist.

The two are set together. Serious firmware, stock or third-party, moves them jointly, with limits, sensors, and sometimes derating. Adjusting only frequency or only voltage because an interface allows it takes the board outside the envelope it was validated in.

Where profiles come from

The vendor publishes modes, with nominal power and hashrate, inside the perimeter of its own testing. Third-party firmware often exposes more sliders, which mostly adds more ways to leave that envelope.

Values that hold on a clean bench do not transfer to a hot hall, a tired PSU, or a different board revision. A profile found online without its air, power, and silicon conditions stays an anecdote rather than a reference.

The six risks

• Instability. Resets, vanishing boards, invalid shares, oscillation between fans and power. The interface can display a hashrate the pool never sees.
• Electrical wear. Higher current on rails, lugs, and PSU. OCP and OTP trips sit closer to their threshold, and a PSU that was barely sufficient on the stock profile no longer is.
• Thermal wear. Hot spots, delamination, accelerated aging of dies and capacitors. Hydro and immersion change the heat path without making an aggressive profile harmless.
• Silent error. A clock too high for the voltage produces bad hashes. The machine hashes, the pool rejects, and local hashrate stays flattering.
• Warranty and site. Off the vendor profile, RMA and the host may refuse coverage. That is a contract clause, checkable before you touch firmware.
• Partial irreversibility. A rail or an ASIC pushed too far does not always return to the behaviour of the gentle profile, even after a rollback.

Underclocking follows the inverse logic: it lowers electrical and thermal stress, and lowers the trial rate too. It is a physical direction to weigh against your electricity cost.

What firmware does not see

Sensors are few, sometimes badly placed, misread, or late to the event. A profile shown green in the interface is not a junction temperature measurement. Custom firmware that ignores a sensor or pins fans to a fixed setpoint adds a risk nothing reports.

The control board can accept a profile the real PSU or the real board will not hold. Behaviour under full load decides, not the configuration screen.

Testing a change without damage

A clock or voltage change modifies the machine's envelope, exactly as a PSU change would. The sequence that limits the damage:

1. Record the original profile and save the firmware image before any modification, so a rollback does not depend on a download.
2. Document the change: who, when, which image, which value, and why. Without that trace, an incident three weeks later cannot be diagnosed.
3. Apply the change to one machine only, keeping its immediate neighbours on the stock profile as a control.
4. Watch three sources in parallel for at least 24 hours: pool-side hashrate, the machine's own sensors, and the temperatures of neighbouring machines in the same rack.
5. Compare pool hashrate against local hashrate. A widening gap means bad hashes, so voltage insufficient for the requested frequency.
6. Roll back to the stock profile as soon as a reset, a missing board, or a rise in rejects appears, before adjusting again.

For official ranges, the only source that binds is the model manual (bitmain.com). A table of frequency and voltage pairs circulating without a vendor document behind it carries no weight on your fleet.