Break even difficulty by hardware efficiency (power cost = value of BTC)

[DeathAndTaxes]

It is useful to know what the electrical break even difficulty of a device in estimating future hashrate growth of the network.  A miner is unlikely to continue to mine for any extended period of time at a negative operating margin.  The break even difficulty is the point where the electricity consumed by the mining equipment equals the value of the coins mined.  At the break even point it will cost the miner the same amount to mine a coin as it would to purchase it.  The hardware cost is not considered in the electrical break even point as once purchased the hardware is a sunk cost and the miner is very likely to continue mining as long as it shows a positive operating margin  (electricity < value of BTC).  Just because the difficulty is below the break even point doesn't guarantee the miner a positive return.

The electrical break even point is based on three factors:
a) the current exchange rate (USD per BTC)
b) the price the miner pays for power (USD per kWh)
c) the efficiency of the mining hardware (J/GH)https://bitcointalk.org/Themes/custom1/images/smflogo.gif

Electrical break even point by hardware efficiency
Based on $100 BTC:USD exchange rate and $0.10 per kWh electrical power rate

Code:

Device        Process  Eff (J/GH) [1]    Diff (mil)  Hashrate (PH/s)
--------------------------------------------------------------------------------------------------
GPU           various       330.0               60              0.4
FPGA          various        50.0              400              2.9
Avalon          110nm         8.8 [6]        2,400             17.2     
ASICMiner       130nm         7.7            2,700             19.3   
BFL (SC)         65nm         5.0            4,200             30.1  
KNC              28nm         1.1 [3]       18,250            130.6   
Bitfury          55nm         0.9           23,300            166.8  
Hashfast         28nm        ~0.8 [7]       25,100            179.7  
Cointerra        28nm        ~0.8 [2]       25,100            179.7   
BFL (Monarch)    28nm        ~0.8 [4]       25,100            179.7  

Update (05/25/2014):  I no longer have the time to research every new ASIC device.   I have not seen a device that is significantly more efficient than 0.8 J/GH (at full clock) when measured at the wall and multiple devices have real world efficiency within +/-20% of that spec so it is a good place to start estimating difficulty. Remember miners are paying for full power consumption not just the power of the raw chip.  The at the wall measurement is the only thing that matters.   More efficient designs may eventually emerge but the low hanging fruit is now gone.  If less than a year efficient of best device went from >8 J/GH to <0.8 J/GH.   We aren't going to see a 1000% improvement in efficient in the next year.  Even a 0.4 J/GH device (at the wall) would be an impressive improvement but only represents a doubling of the break even point.

To calculate the break even difficulty at a different exchange and/or power rate use the following formula:

Code:

Adjusted difficulty = base difficulty * (exchange rate / $100) * ($0.10/power rate)

Examples:

Code:

Avalon break even difficulty @ $200 exchange rate and same power rate.       2,400 * ($200/$100) * ($0.10/$0.10) =  4,800
BFL (SC) break even difficulty @ $0.15 power rate and same exchange rate.    4,200 * ($100/$100) * ($0.10/$0.15) =  2,800
Bitfury break even difficulty @ $200 exchange rate and $0.05 power rate.    21,000 * ($200/$100) * ($0.10/$0.05) = 84,000

Why J not W?
Because J (Joule) is a measure of energy and W (Watt) is a measure of power.  One Joule per second is one watt.  You can't compare watts (energy over time) to a timeless value like Gigahash.  Efficiency can be measured either in W per GH/s or J/GH bu not W/GH.  It would be like saying a car gets 30 MPH per gallon or the spot price of gold is $1500 per ounce second.  You can safely assume is someone writes W/GH they mean J/GH (or W per GH/s).  I put this here because I have already gotten "corrections" by PM.  Simple version: W per GH/s or J/GH are correct, W/GH is not.


Why did you make this?
I find it silly people have projections with difficulty going to 1 trillion or more (I think the highest I have seen is 200 trillion). At a mere 300 billion difficulty all current and proposed mining devices would be operated at a significant loss (assuming $100 exchange rate & $0.10 electrical rate).  So at 1 trillion in difficulty even a Cointera rig would be converting $4 in electricity into $1 in Bitcoins.  Anyone think that is likely?   Significantly higher difficulty is going to require either more efficient hardware, a higher exchange rate, or the average cost of power for the network to decline.


Notes:

[1]  Numbers used are from mining hardware comparison where available or based on benchmarks reported by owners/manufacturer.  For apples to apples comparison all units are based on wattage at the wall (including all system power and inefficiencies).   For devices where total wattage at the wall is unavailable it has been estimated.   For unreleased devices ("specs") the efficiency reported by manufacturer is used.  

[2] Cointerra did not provide system efficiency "at the wall" however did indicate that chip wattage would be better than 0.55 J/GH, based on that I project the at wall efficiency to be ~0.7 J/GH.  Updated to 0.6 J/GH at the chip and 0.75 J/GH at the wall based on official statement.  This is based on the following assumptions.  Chip power: 0.60 W per GH/s * 2000 GH/s = 1200W @ 0.785 VDC.  DC to DC board PSU 90% efficient 1200W/0.9 = 1333W @ 12VDC.  Balance of system controller & cooling/fans = 50W @ 12 VDC.  Total system DC wattage 1383W @ 12VDC.  ATX PSU 90% efficient 13832/0.9 = 1537W @ 120VAC.  1537W / 2000 GH/s = ~0.8 J/GH.

[3] Updated to 1.1 J/GH (https://bitcointalk.org/index.php?topic=170332.msg3307091#msg3307091). 455W @ 1 VDC running at 502 GH/s.  Assume 90% DC to DC conversion and 90% AC power supply efficiency results in estimated 560W at the wall.    

[4] The Monarch card is reported by BFL to use 350W.  It is not clearly indicated if this is 350W at the chip or the card but I will assume the 350W is the total 12VDC wattage for entire card (to include one or more ASIC processors, host interface, card fan, and DC to DC board PSU.  The balance of system is provided by the user so some assumptions are necessary to estimate efficiency at the wall.  To be used in a PCIe slot the card requires a "traditional motherboard" with one or more PCIe slot, 100W min is assumed for system and fans.  Total DC wattage for single card system is 450W DC.  Assuming a 90% efficient ATX PSU 450W DC/0.9 = 500W @ 120V AC.  500W/600GH/s = ~0.8 J/GH.  Actual wattage may be significantly higher depending on the components selected by the user.  

[5] Given BFL's delivery track record and after the fact changes I don't feel comfortable making any prediction.  I would recommend users be very conservative and pad the delivery timeframes provided by BFL.

[6] Due to multiple configurations power efficiency varies by number of modules, batch version, and PSU efficiency.  The following efficiencies have been reported 9.4 J/GH (66 GH/s @ 620W), 8.8 J/GH (85 GH/s @ 750W), 8.5 J/GH (82 GH/s @ 700W).

[7] Based on reported efficiency at the wall of less than 350W and nominal performance of greater than 400 GH/s of the BabyJet with single hashing board.  

[1715366520]

1715366520

[1715366520]

1715366520

[1715366520]

1715366520

[1715366520]

1715366520

[fpgaminer]

$0.30+ per kWh reporting in 

[k9quaint]

$0.30+ per kWh reporting in 

Hello fellow Californian?

[DeathAndTaxes]

$0.30+ per kWh reporting in 

Ouch.  The break even difficulty and hashrate would be 1/3rd of what is listed. 

[eraziel]

Do I read this correctly, GPU miners are already not getting enough btc to break even?

[DeathAndTaxes]

Do I read this correctly, GPU miners are already not getting enough btc to break even?

At $100 exchange rate and $0.10 electrical rate?  That is correct.  Obviously each miner's efficiency will vary as will their electrical rates and their "pain threshold" (how negative of a margin they are willing to accept before shutting down) but generally speaking yes.

[cryptx]

Cointerra has the most power efficient 28nm process at the moment.

1666 MH/J

[Dexter770221]

Cointerra has the most power efficient 28nm process at the moment.

1666 MH/J

Really?
They demonstrated that? It's for whole platform (like in OP charts) or chip itself? From 1666 you may easily get down to 1200 with poor PSU design...

[eraziel]

Cointerra has the most power efficient 28nm process at the moment.

1666 MH/J

Really?
They demonstrated that? It's for whole platform (like in OP charts) or chip itself? From 1666 you may easily get down to 1200 with poor PSU design...

Most efficient vaporware indeed
I hate it when people start stating claims based on marketing materials as if they are a fact...

[cryptx]

Really?
They demonstrated that? It's for whole platform (like in OP charts) or chip itself? From 1666 you may easily get down to 1200 with poor PSU design...

I guess with KNC and Hashfast also in the list, you have to be consistent and also include Cointerra.

OR

You exclude all chips currently under development.

[BurtW]

I believe this thread is trying to calculate the "theoretical maximum economically sustainable system hash rate given various technologies", right?

Therefore all vaporware should simply be included in a separate category:

Code:

Specs     Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s)
BFL          28nm ...
Cointerra    28nm ...
Hashfast     28nm        1200      25,146        180,000
KNC          28nm         400       8,382         60,000

Device    Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s)
Avalon      130nm         120       2,515         18,000 
ASICMiner   110nm         130       2,724         19,500 
BFL          65nm         200       4,191         30,000 
Bitfury      55nm        1000      20,955        150,000 
        
FPGA      various          20         419          3,000
GPU       various           3          63            450

BTW, Thanks!  And this is very interesting.

[AsicShill]

It certainly puts a feather in the cap for Bitfury.  Of the shipping chips they are way out ahead of the pack (by 5x!). 
Though at current pricing of $35 -> $54 per Gigahash (depending on overclock), the pricing is a bit meh.

https://megabigpower.com/shop/index.php?route=product/product&path=60&product_id=62

[VinceSamios]

Anybody care to add a calculation for the total cost to achieve the listed hashrate for each of the respective models?

ie.
for the HashFast product and corresponding breakeven point of 25trillion it's $352Million in hardware costs
For BFL's 65nm is $188Million

etc.

[cryptx]

Code:

Specs     Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s) Cost(mil)
BFL          28nm        1666      34,925        250,000      1,950
Cointerra    28nm        1666      34,925        250,000      1,969
Hashfast     28nm        1200      25,146        180,000      2,520
KNC          28nm         400       8,382         60,000      1,050

Device    Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s)
Avalon      130nm         120       2,515         18,000 
ASICMiner   110nm         130       2,724         19,500 
BFL          65nm         200       4,191         30,000 
Bitfury      55nm        1000      20,955        150,000 
        
FPGA      various          20         419          3,000
GPU       various           3          63            450

[VinceSamios]

Here I've added cost of the cheapest technology to make this particular technology obsolete

Code:

Specs     Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s) Cost(mil)      Minimum Replacement Cost(mil)
BFL          28nm        1666      34,925        250,000      1,950      1,950
Cointerra    28nm        1666      34,925        250,000      1,969      1,950
Hashfast     28nm        1200      25,146        180,000      2,520      1,440
KNC          28nm         400       8,382         60,000      1,050      480

Device    Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s) Cost(mil)
Avalon      130nm         120       2,515         18,000      2,520      144
ASICMiner   110nm         130       2,724         19,500      1,872      156
BFL          65nm         200       4,191         30,000      1,410      240
Bitfury      55nm        1000      20,955        150,000      3,000      1,200
        
FPGA      various          20         419          3,000
GPU       various           3          63            450

[DeathAndTaxes]

Cointerra has the most power efficient 28nm process at the moment.

1666 MH/J

Do you have a link to this.  I thought Cointerra simply said significantly below 1GH/W.  Also I need full system power not the chip.  Even if it is an estimate the manufacturer I will use it if they are willing to put it into writing.

[DeathAndTaxes]

Code:

Specs     Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s) Cost(mil)
BFL          28nm        1666      34,925        250,000      1,950
Cointerra    28nm        1666      34,925        250,000      1,969
Hashfast     28nm        1200      25,146        180,000      2,520
KNC          28nm         400       8,382         60,000      1,050

Device    Process  Eff (MH/J)  Diff (mil) Hashrate (TH/s)
Avalon      130nm         120       2,515         18,000 
ASICMiner   110nm         130       2,724         19,500 
BFL          65nm         200       4,191         30,000 
Bitfury      55nm        1000      20,955        150,000 
        
FPGA      various          20         419          3,000
GPU       various           3          63            450

The issue here is you are comparing the chip level for one product to the board level for another product to the system level for the rest of the products.

For example if BFL delivers a 600 GH/s card which has 1,666 MH/J efficiency it requires a host system with one or more PCIe slots (50W min), case fans (3x12W ea?), and a AC PSU which is say 90% efficient at load.  So BFL would have a total DC load of 436W which is a 484W AC load (assume 90% efficient PSU).  600 GH/s / 484W = 1240 MH/J.  Of course with more fans, higher system wattage, or worse PSU it would be even lower.

Cointerra is even more an unknown.  We don't even know the board level wattage (@12V), what type of controller will be used, is it a custom enclosure, if so how many fans, etc.  That makes trying to estimate a system wattage totally unknown.  Also as indicated above I can't see to find a cite for Cointerra indicating a specific efficiency even just for the chip. However even if Cointerra chip(s) deliver 2TH/s using 1200W the ASIC (like all other ASICs will run at @0.8V to 1.2V) and need a DC to DC PSU to convert the 12V supplied by the AC PSU to the ~1V required by the chip.  The DC wattage would then be 1200/0.9 = 1,333W @ 12VDC.  If we assume the balance of the system is a Pi or other micro controller (5W and 3x 1A fans) it would make total system wattage 1374W DC = 1527W AC.  That would put the efficiency at the plug ~1300 MH/J.

Simple version
System AC ("at the wall") wattage = System DC Wattage / ATX PSU efficiency *   
System DC wattage = ASIC Board(s) wattage + balance of system (controllers, fans, waterpump, etc) **
Board DC Wattage = ASIC Wattage / DC to DC PSU ***

* Depends on ATX PSU efficiency at the particular load.  As a power customer you pay for wattage at the wall.  This will always be higher than the DC wattage used by the system due to PSU inefficiency.  If unknown a good upper bound is 90% efficiency.  More efficient PSU are possible but they generally require 230V and are more expensive (google "80 Plus Titanium")

** Rasberry Pi uses <5W.  BFL PCIe card will require a system with one or more PCIe slots which means a "traditional" motherboard and >50W.  If case fans or pumps are unknown an estimate of 12W per 120mm fan and 20W per pump (HashFast) is reasonable.  Without knowing the exact components we can't know the exact system wattage but for example nobody is going to use fans which use half a watt or a water pump that uses 100W or PC host which uses 10W.

*** ATX PSU supplies high current power at 12V however no ASIC runs at that voltage.  Each ASIC board will need to convert the 12V to the voltage used by the ASIC.  This will vary from ASIC to ASIC but is generally in the range of 0.6V to 1.3V.  Good high current DC to DC power supplies are expensive (up to $1 per watt) and generally are less than 90% efficient.  Lower cost PSU will provide significantly reduced efficiency (80% to 84%).

[cryptx]

Cointerra has the most power efficient 28nm process at the moment.

1666 MH/J

Do you have a link to this.  I thought Cointerra simply said significantly below 1GH/W.  Also I need full system power not the chip.  Even if it is an estimate the manufacturer I will use it if they are willing to put it into writing.

You can find it here:
http://thegenesisblock.com/cointerra-announces-2ths-asic-bitcoin-miner-for-15750/

[BurtW]

So it looks pretty safe to say that the network will be less than 250 PH/s using the 28nm technology at the current prices.

[DeathAndTaxes]

So it looks pretty safe to say that the network will be less than 250 PH/s using the 28nm technology at the current prices.

The replacement cost is calculated based on all hashing power coming from a single design so it shouldn't be taken as likely but more an upper bound.  I think there might be an error in the math as well.

The cost to reach the "break even point" is based on current cost so I don't find it too useful, because cost will decline significantly.  The network will approach a small margin below the break even point.  It is only a matter of time.  You can consider that the equilibrium point, we have seen it occur in the GPU era for a couple of years.  If more efficient hardware is produced then the hashrate will rise, if the exchange rate rises then the hashrate will rise, as mining moves to lower cost areas the hashrate will rise, if the exchange rate falls the hashrate will fall.  However the network will approach and stay close to a margin below the break even point.  How close?  It really depends on miners and what risk they are taking.