It is a quiet evening and I am enjoying a large cup of coffee while thinking about GPU mining efficiency. The way I see it, there are two kinds of maximum efficiency, and it is not completely clear to me yet which one is the ‘best’ kind.
On my present video card (a 2-year old MSI GeForce® GTX970) mining algorithms like Equihash seem GPU-restricted. I know crypto currency mining is supposed to require fast memory but what I see is always 100% GPU load while memory controller load will float anywhere between 65% and 85% depending on the selected power limit and GDDR5 clock. Heck, I can even remove the +500 MHz memory overclock without influencing calculation performance.
Anyway, all microprocessors have a power/performance curve which is linear for a time and then has a knee point where e.g. 10% more performance requires 25% more power. On the other hand, substantial reductions of power limits bring less significant reductions in performance.
All video cards are designed to operate somewhere below that knee point and have a 10-20% power limit reserve before they start to throttle down due to insufficient voltage stability or overheating.
The first kind of mining efficiency, the ‘mine all the coin you can’ kind, relies on you finding that point where the GPU consumes maximum power without going over the temp limit (throttling) or crashing when rushed with extra work (if overclocked without sufficient voltage boost…)
However this kind of maximum yield efficiency might not be reasonable for me because it wears down the equipment very quickly, causes a lot of noise that makes my cats nervous and increases the electricity bill.
In the case of the GTX970, top power consumption (109% power limit) results in efficiency of 1.2~1.3 H/W. If I dial down power limit to 65-75%, I can nearly double the efficiency to 2.1~2.2 H/W which keeps my card at around 60° and presumably makes polar bears happier.
From economy standpoint, this best power efficiency seems optimal: you mine the largest amount of coin per Joule. Your equipment will last much longer, and your energy bill will be lower.
Power consumption is always important, because it is a fixed cost that changes linearly with the power limit of the card. But achieving best power efficiency also reduces overall performance: at 109% I can achieve ~260 MH/s Sols/s, whereas at 70% I get ~190 MH/s.
This translates into a very palpable difference in net daily profit. Over the course of one day the lower rate would yield ~4.90 BGN/day whereas the higher rate would yield ~6.20 BGN/day: a delta of 25% or a whole 1.30 BGN/day more.
At the same time, my energy cost will rise proportionally but not that significantly in net values.
My cost of energy per kWh depends on the time of day. We have a daily rate of 0.22 BGN/kWh (~0.11 EUR) which runs for 16 hours and a nightly rate of 0.13 BGN/kWh (~0.066 EUR) which runs for 8 hours.
Electricity cost/day where I live:
70% power limit (~100W): 1.6 kWh × 0.22 + 0.8 kWh × 0.13 = 0.46 BGN
109% power limit (~160W): 2.56 kWh × 0.22 + 1.28 kWh × 0.13 = 0.73 BGN
The difference between power limited and full power brings savings of 0.27 BGN/day but leaves 1.20 BGN of mining profit on the table. I am not willing to accept that waste.
Of course what is missing from this evaluation is the increased potential risk of damaging my hardware or drastically shortening its life. I may yet be proven wrong about this if my card dies, but for the time being I will choose maximum yield efficiency over best power efficiency.
Note: I was a little lazy with these calculations as I was taking monetary values directly from the NH miner app. This is generally not a good idea because fluctuations of the price of BTC will also affect the monetary estimate; a better solution is to monitor the BTC daily yield figure. But I checked and made sure that the BTC/BGN was ratio was quite stable (about 1% corridor) on the day I wrote down these numbers, so they are fairly representative. But for goodness sake don’t ever try this when there are big spikes in BTC price — these will make you blind to actual figures.