Insights
Jun 12, 2024
Introduction
This whitepaper introduces a new Bitcoin mining management software, Hashman, which optimizes ASIC power modes for maximum profitability.
We demonstrate >18% profit improvements from real-life data over simpler on/off automation, when only the average ASIC fleet efficiency is considered, for a single model of Bitcoin miner (here, Antminer S19j Pros).
Overview of Hashman
Hashman is designed from the ground up to minimize human operator time, while boosting profits near to theoretical maxima.
In contrast to previous publicly known approaches, Hashman learns the individual characteristics of each ASIC, and automates operations taking into account the input costs and revenues, i.e. static or market-priced electricity costs, and approximate hash prices from Bitcoin spot prices, transaction fees, and the block difficulty level.
Taking into account revenues deriving from the hash price, and costs from energy usage, Hashman turns on all ASICs which yield non-zero profit, and off anything that does not have a power mode that can yield profits, given the energy costs.
Further, given that ASICs take a non-negligible amount of time to switch between power modes, we create a power schedule for up to 48 hours in advance (recomputed on input cost/revenue changes), which can minimize superfluous power mode transitions and further boost uptime.
Case study: >18% profit boost in challenging backtest conditions over typical baseline conditions
We provide a case study using a fleet of Antminer S19j Pros running BraiinsOS with autotuning. Despite running an apparently "identical" set of machines, we can still get sizable gains from applying a bit of intelligence, as shown below. We use autotuning to get a baseline boost for, while obtaining fairly accurate individual power usage and efficiency data with which to calculate optimal power schedules per individual ASIC.
Note: Profitability gains are expected to be larger with heterogeneous fleets where there is greater variability.
We choose 2022-01-01 through 2024-04-20 as the range for backtesting, and we believe that much earlier data does not sufficiently reflect the present day.
To provide a fair approximation of the future, we assume that the hash price (sat/PHs/day, or sat/EH as we like to denote it) was half of the real values before the fourth halving. Otherwise the backtest reflects real historical data.
Fleet data
Model: Antminer S19j Pro 104 TH/s
Firmware: BraiinsOS 23.08 and above
Count: 446 distinct ASICs
Power modes under auto-tuning with power targets:
1700W (lowest supported setting)
2400W (approx. mid-point between low and high)
3100W (approx. the default power usage for this model)
Note: As per other backtests and as recommended by Hashman, using more than 3 power modes provides no additional benefit.Instead, overall profitability is maximized by increasing the absolute range of power targets.
We measure the following as the average efficiencies for these ASICs with varying ambient temperatures ranging from ~0°C to ~20°C, while mean operational chip temperatures hover around 60..70°C. We impute these as assumptions in the backtest without taking into account any variability-over-time due to environmental factors or otherwise (in contrast, Hashman does constantly adapt the efficiency estimates to take into account such changes)
Power target = 1700 W: Average efficiency = 25.00 J/TH; average power usage = 1642 W; average hashrate = 65.69 TH/s
Power target = 2400 W: Average efficiency = 27.74 J/TH; average power usage = 2399 W; average hashrate = 86.48 TH/s
Power target = 3100 W: Average efficiency = 30.10 J/TH; average power usage = 3073 W; average hashrate = 102.10 TH/s
Electricity prices
We use Nordpool/Finland spot electricity prices as the data source for energy costs, and add VAT charges to simulate more realistic usage.
Scenario 1: Average fleet efficiency, single power mode
For a baseline scenario, we consider optimal automation with the following assumptions:
The entire fleet is operated as a single unit
Only a single power mode (i.e. power target) is used. For this data set, 3100 W maximizes profits (i.e. ASICs will switch between this and "sleep mode")
ASICs are turned on and off depending on the average parameters of the fleet (as above)
Note: While operating based on the average characteristics is not optimal, assuming the data is accurate, it is still profitable.
Scenario 2: Average fleet efficiency, multiple power modes
The second scenario improves upon the first:
The fleet is still considered a single unit that turns on and off together
However, we use the three power targets from above, and do so optimally. I.e. for each time period, the power target is chosen which maximizes the profit for that time (this may be "sleep mode" at times).
By selecting from low/mid/high power targets, we increase the profits.
Scenario 3: Individual ASIC efficiencies, multiple power modes
The final scenario brings to the fore (we think) the maximum theoretical efficiency from ASICs by using knowledge of the individual ASIC variations to individually construct a power schedule.
The fleet is just a set of ASICs that operate individually
Hashman operates each ASIC per its power schedule
Each power schedule is optimized taking into account power mode transition times, only allow downtime from a change if the total profit over the planning horizon (e.g. 24 hours) exceeds the alternative of keeping another power mode on for longer.
Conclusion
Aside from operational improvements due to more reliable automation, Hashman can deliver substantial bottom line improvements to ASIC profitability by learning each the characteristics of each individual ASIC and using the individual device as the unit of optimization, rather than an entire fleet, or a set of "identical" machines, like the Antminers in this case study.
