ASIC Efficiency

In mining, each hash calculation is another attempt to find a result that meets the requirement for a new Bitcoin block. Hashrate measures the speed of these attempts. TH/s means terahashes per second; one terahash represents a trillion calculations. Power draw, meanwhile, tells us how quickly the equipment consumes energy. A joule is a unit of energy. Divide electrical power in watts by Hashrate in TH/s. An illustrative machine drawing 4000 W at 200 TH/s gives 20 J/TH. The seconds cancel because a watt is a joule per second. The result tells us the energy cost of the same amount of computational work. [1] [5]

Consider a real document rather than a current purchase offer. The S21 Pro specification from 2024 lists 234 TH/s, a power draw of 3510 W and 15 J/TH at an inlet air temperature of 25 °C. The manufacturer calls these typical figures and specifies tolerances. Eva cannot therefore expect every unit to produce exactly the same result at any temperature. A specification is a starting point for comparison; a number without its measurement conditions could create false certainty. [1]

If Eva wants to understand operating bills, a chip-only figure is insufficient. Measurement at the whole device’s electrical input includes everything powered through that input, including the corresponding losses. A separate pump or building cooling system may use another supply, however, so its consumption must be added separately. Braiins recommends actual measurement because software power estimates may be inaccurate. Comparing a chip’s efficiency with that of an entire facility would mean comparing different things. [2]

Manufacturing differences between chips and temperature affect operating results. Braiins explains that the same tuning process can yield different improvements in different machines. Eva therefore records temperature, operating mode and firmware, the software controlling the equipment, alongside her measurement. Power draw and Hashrate must cover comparable periods. A brief reading during startup or tuning cannot automatically be treated as the stable performance the machine will sustain in ordinary operation. [3]

Autotuning automatically searches for suitable voltage and operating frequency for individual chips. Frequency determines the pace of their work. Overclocking, operating above the usual working frequency, can increase Hashrate while worsening J/TH. A higher hashrate alone therefore does not establish better energy efficiency. Eva must examine both quantities and the actual outcome, not merely the limit set in a dashboard. Equipment capabilities and power-supply or cooling constraints still apply. [2]

Take a purely illustrative change: instead of 200 TH/s at 4000 W, a machine runs at 150 TH/s and 2700 W. Its figure improves from 20 to 18 J/TH. It uses 10% less energy for the same work, but performs a quarter fewer calculations in the same time. Those are different changes. Lower power draw may suit Eva, but she must also consider the reduction in expected revenue. These figures are not recommended settings for a particular machine. [2]

For several machines, simply averaging their J/TH figures is insufficient. A faster machine contributes more to the group’s work. Imagine an illustrative 200 TH/s at 4000 W and another machine providing 100 TH/s at 3000 W. Add both power draws and both hashrates: 7000/300 gives roughly 23.33 J/TH, not the simple average of 25 J/TH. Measurements must cover comparable conditions and periods. This captures energy per work for the group rather than giving a small and a large machine equal weight. [1]

Alongside efficiency, Eva needs Hashprice, the expected gross revenue per unit of hashrate over a specified period, electricity prices, actual operating time and fees. Purchase price and other costs also matter for payback. More efficient equipment can cost more, and J/TH alone cannot settle the choice. Total consumption need not fall either: if Eva adds machines or raises their output, the whole operation can draw more power even while using less energy per calculation. [4]