Underclocking runs a chip at a lower operating frequency than a chosen reference setting, such as the manufacturer’s default profile. It can help reduce power consumption and heat. By itself, it does not mean lower voltage, better energy efficiency or a complete halt to mining.
An ASIC is a specialized chip that performs mining calculations. Its operating frequency, expressed for example in MHz, sets the pace of that work. MHz means megahertz: millions of clock cycles per second. These cycles are not the same as completed hashing attempts. Paul first needs to identify the profile he is lowering it from: lower means little without a reference. Braiins explains that slower operation may also allow lower voltage. Changing frequency alone, however, does not guarantee a voltage change. Mining Firmware, the software controlling the machine, may adjust both together; they remain different adjustments. [1] [6] [7]
Imagine the chip performing fewer working steps per second. The energy associated with its switching activity may fall. Braiins simplifies the relationship as W = a × F × V²: Here W denotes power draw in watts, F frequency and V voltage. The letter a is a coefficient for the model. Voltage is squared in this model. This explains a relationship rather than precisely calculating the whole mining system’s consumption. Fans also draw electricity, and the power supply has losses. Paul therefore cannot read an exact electricity bill from a frequency setting. [1] [7]
Consider an illustrative example: 2000 W / 100 TH/s = 20 J/TH. After a change, we measure 1200 W / 50 TH/s = 24 J/TH. TH/s means terahashes per second, or trillions of hashing attempts per second. J/TH means joules per terahash: the energy needed for a trillion such attempts. Power consumption fell by 40%, but the new mode uses 20% more energy for the same number of attempts. Reducing circuit load may be useful, yet energy efficiency has worsened. These are illustrative figures, not recommended settings for a particular machine. Paul needs to measure both consumption and work performed. [1] [5] [6] [7]
When Paul changes a setting, an immediate acknowledgement confirms receipt, not necessarily achievement of the target. LuxOS documents a gradual transition for frequencyset. He can check current frequency with frequencyget or use IsRamping to see whether a transition is underway. The S9 also supports only certain frequency steps. The returned Frequency field therefore states the actual available target, the closest supported value, which need not exactly match the request. That target alone still does not prove that the chips are already operating at it. [2]
Paul need not initiate every change manually. LuxOS uses ATM, Advanced Thermal Management, to regulate operation according to temperature. It can reduce frequency as the machine heats up and raise it within permitted limits after cooling. All available temperature inputs must permit an increase, not just one cooler board. Depending on the hardware, these include chip temperatures and, for water-cooled equipment, inlet water temperature. The reasoning is straightforward: one cooler component does not mean the whole machine is safely cool. [3]
Underclocking cannot guarantee uninterrupted operation with arbitrarily poor cooling. In LuxOS, Minimum Profile sets the lowest profile ATM may use; power_min supplies the lower boundary in power-targeting mode. If Paul sets that floor too high and overheating continues, regulation has nowhere further to retreat. Firmware may then shut the device down. The floor is not a promise that the machine must keep producing at least that much under every condition. Thermal protection takes priority over the desired performance. [3]
Braiins uses DPS, Dynamic Performance Scaling, to change power or hashrate targets as conditions change. Hashrate means hashing attempts per second. Returning to a higher level depends on version, mode and timing conditions. DPS Memory accounts for earlier machine responses to prevent rapid switching back and forth. Lower ASIC performance therefore need not mean a fault, even when the interface offers a higher profile. Paul first checks temperatures and logs for the reason. Disabling protection just to raise the displayed number does not remove that cause. [4]
Underclocking can keep mining running at a slower pace. Pausing calculations completely is a different state and need not mean zero power consumption either: control electronics or cooling may still need electricity. Paul compares equal periods under similar temperatures, whole-system wall power, errors and work accepted by the pool, a service coordinating miners. A software power estimate is not an independent measurement. Together these observations reveal whether he met his goal. Slowing down alone promises neither savings, longer service life nor profit. [1] [5]
For the clearest picture, read this entry together with ASIC Miner, Autotuning, Mining Firmware, ASIC Efficiency. The reverse links also lead from Autotuning, Overclocking.