http://toom.im/images/Spondoolies%20Hashrate.png
http://toom.im/images/Spondoolies%20Efficiency.png
http://toom.im/images/Spondoolies%20Marginal%20Hashrate.png
The marginal efficiency graph shows the efficiency of a small increment in power. If you step the power consumption on an SP20 up from 1100W to 1200W, you only gain an extra 50 or so GH/s for 100 W in extra power consumption, giving a marginal efficiency of 2.0 J/GH. On the other hand, the overall efficiency at 1200W for an SP20 is about 0.75 J/GH.
The marginal efficiency is a very useful metric for choosing what power settings to run a miner at. You should run your miners at a setting such that the marginal efficiency leaves you at the breakeven point. Right now, 1 TH/s earns about $0.111 per hour. If your power or hosting costs you $0.14/kWh, then your breakeven point would be
[$0.111/(h•TH/s)] / [$0.14/kW•h] = 0.79 kJ/TH = 0.79 J/GH
so you would want to run an SP20 at the setting that gives you a marginal efficiency of about 0.79 J/GH, which turns out to be about 600 W AC.
Some notes:
All tests were conducted on 240V +/- 5V with intake temperatures between 8°C and 18°C.
The SP31s tested used Murata power supplies. The Murata PSUs are more efficient than the Emerson1200 PSUs (about 91.5% vs 88%). The SP30 tested used Emerson1200s. The SP20 used an HP DPS1200FB PSU, which appears to be 90% efficient on 240V.
AC power consumption for the SP35, SP31, SP10 and SP20 were measured using a Dell PDU, and should be accurate to within 20W. AC power consumption for the SP30 was read off of the ASIC Stats page. Previous testing has shown the Emerson1200 and Emerson1600 (in the SP35) to give accurate AC consumption measurements compared to our PDUs.
The DC power measurements of the SP10s and SP20s appear to be miscalibrated. At high power, they appear to be accurate, with our PDU measuring AC consumption that's about 10% higher than the reported DC power consumption. At low power (e.g. 60%), the SP10 reports DC power consumption 4% *below* the AC consumption, and the SP20 reports DC consumption equal to AC consumption. Power measurements from Murata power supplies show an inverse pattern, and are very inaccurate (and biased for low readings) at full power output, but tend to be accurate at 80% of max power or below. Do not rely on the power readings from SP10s, SP20s, and Murata SP30/31s for anything important.
The SP10s tested were all from a moderately slow corner batch.
A few of the max power datapoints might be missing from some machine types (e.g. SP10, SP35) to simplify testing. Most testing was done using voltage limiting (600 mV < start voltage = max voltage < 700 mV, or 750 mV for SP20). The SP30 was tested using power limiting, which is part of why its graph is slightly wavy.
At very low voltages or power limits (< 620 mV, roughly, for both Rockerbox and Hammer), some ASICs on a machine will fail to start up. This results in rapidly dropping hashrate and power consumption towards the left end of the graphs, as well as potential system instability, with slight improvements in efficiency. Different ASICs will have vastly differing minimum voltages for operation, which means that some machines will be capable of much lower power operation when run at minimum voltage. I've seen a 25% (430W) difference in minimum power between two SP31s from the same batch, for example. However, the difference in efficiency between these machines at their minima was smaller, at 0.50 vs 0.45 J/GH. YMMV (your miners may vary).
Data tables available upon request. Email [email protected] if you want one. Specify which machine type(s) you want them for.
