ASICMINER: Entering the Future of ASIC Mining by Inventing It

[xhomerx10]

 You have begun with the incorrect assumption that your bitcoin miner will hash at a constant percentage of the network hashing rate for 730 days.  I fail to see the usefulness of your graphs.

It makes no such assumption. As long as the network hashrate is at or below the lines on the chart, at any point in time you could buy a miner and end up breaking even after 2 years, of course constrained by the listed assumptions and the curves themselves, nothing else.  If your point is that after 2 years the network may exceed those lines, thats actually part of the point, but it  requires either changing constraints or someone betting > 2 year.

 Let's dissect the data where all points converge @1239PH/s network speed and no cost for electricity.

1 TH/s is costing $700 now on your chart and bitcoin is worth $330;

 1239 PH/s = 1239000 TH/s  
Take the reciprocal of this 1/1239000 to find out your share of the block rewards for 1 TH/s and multiply that by the total number of bitcoins produced in a day.

 total bitcoins per day for the entire network (ideal):
 25 per block (reward) x 6 (blocks per hour) x 24 hours per day = 3600

 3600 multiplied by 1/1239000 = 0.0029055690072639225181598062954 Bitcoins per day (ideal)

Now if you make that every single day for the magical number of 730 days you get 2.121065375302663438256658595642 bitcoins
and when we multiply that by the assumed $330 per bitcoin... drumroll please... we get  $699.95157384987893462469733656186 or approximately $700 dollars for our 1 TH/s miner that we paid $700 dollars for 730 days ago!  Which only works for a sustained network hashing rate of 1239 PH/s for a period of 730 days.

 Your graph and your assumptions are flawed.

           ^that is my point

The assumptions are just that, feel free to alter them yourself. The graph isnt flawed, you just dont understand what its telling you, even after manually verifying the math.

What you fail to understand is that the network isnt going to keep growing magically if there is no ROI to be had. That graph tells you when there is no more ROI to be had (or at least not within <2 year, which seems a reasonable horizon to me but feel free to change that too). Once you reach that point, who is going to buy or deploy more miners? Few if anyone, and if someone does, others will have to unplug, hence the network hashrate will remain roughly where the chart tells you it will plateau (for your chosen assumptions).

 First of all, it's not my math, it's your math.  You said you did not make the assumption that the network will remain at a constant rate for two years.  Your math shows you did but whether you did so consciously or not is an altogether different argument.

 Secondly, your charts have nothing to do with network growth as you have only made very narrow and fixed assumptions.  So my failure to understand network growth based on your primitive charts is not at issue here.  

[Puppet]

Dude you are thickheaded. Its not a tool to help you predict if you will make a profit from a miner, its tool that predicts where the network hashrate is ultimately headed.

[teek]

 First of all, it's not my math, it's your math.  You said you did not make the assumption that the network will remain at a constant rate for two years.  Your math shows you did but whether you did so consciously or not is an altogether different argument.

 Secondly, your charts have nothing to do with network growth as you have only made very narrow and fixed assumptions.  So my failure to understand network growth based on your primitive charts is not at issue here.  

I think the chart does look right to me..  it's also pretty interesting.. 

He didn't make the assumption the network will remain at a constant rate for two years,  he is just using 2 years as an example and extrapolating the data over a arbitrary period of 2 years.  He could have picked 6 months or 5 years..  The curves may be condensed or elongated in reality..  no one knows exactly how the network will play out.   Price also plays into this big time as the final cost of electricity is relative..  Put btc at 10k and this chart goes on forever..   

This is about at what network hashrate versus  mining efficiency the entire network has to stop growing..   

For example:  Say in our current situation..  AM gear is at about 0.7 (the orange line)  assuming an electricity price of 0.06KWH,  the gear is unprofitable at the point which the network reaches 850 PH,  WHENEVER that may be,  be it 3 months from now or 2 years from now.   Ie:  If the network was suddenly 850ph tmrw,  the gear is no longer profitable.  On the other hand,  if the network took 5 years to get to 850PH (again forgetting about the block halving / price differences)  the gear would still be profitable until that time..

However gear of greater efficiency say 0.1,  could hang around until 1850PH (still assuming 0.06kwh)  still being profitable.. 

Finally,  if we were to say that mining gear hits an efficiency floor at 0.1 J/GH and electricity is 0.02KWH (there is no such thing as FREE energy) the network will stop growing at approx 2300 PH.

Only the price of Bitcoin (and thus the relative electricty cost) is going to affect these numbers.. 

teek


 

[xhomerx10]

Dude you are thickheaded. Its not a tool to help you predict if you will make a profit from a miner, its tool that predicts where the network hashrate is ultimately headed.

 That's not what it showed in your first post.

Doesn't matter what the J/GH is if you can never recover the capital cost of miners even with free electricity, which is where bitcoin mining stands right now.

You couldnt be more wrong. Here is a chart for you:



It shows the network speed where miners would break even after 2 years using the listed assumed variables. Even  in the current climate and with current efficiency, we are no were near where (industrial) mining  would not be profitable. And the effect of power efficiency is quite dramatic if you consider reasonable electricity cost price ranges (~0.06 / KWh)

 ...oh and dude?  It shows neither.

[Puppet]

THe chart was posted for those smart enough to understand what it means; if you're not among those, feel free to ignore it, just like I will be ignoring you.

[ensurance982]

Huh, funny... as if the price can't quite decide which way to go! Strange... people selling huge volume down, yet people are buying at a rather decent spread...

[jjdub7]

THe chart was posted for those smart enough to understand what it means; if you're not among those, feel free to ignore it, just like I will be ignoring you.

I've actually produced similar outputs back in some engineering classes for non-linear, non-isentropic processes (turbine propulsion), and cross-sectional/contour data is great for communicating data like this.

While the data indeed communicates the engineering/design side of the equation, one assumption that I think many people forget is that depreciation (i.e. resale value) on mining equipment has historically progressed at a relatively moderate pace (though granted, liquidity of the resale market might dry up quickly and without advance notice).  Essentially, from my perspective on the secondary hardware markets, buyers of used equipment almost seem to be willing to pay the time-discounted equivalent of what the MSRP would be on comparable but broader markets for consumer electronics (e.g. GPUs, which are more general-purpose).  I'm guessing this is due to the fact that the supply of ASIC mining equipment still remains relatively scarce due to the niche demand for such devices.  That being said - the highest resale values by model year/process architecture/"snapshot generation" of equipment are always going to go to the most efficient equipment (with considerations such as interchangable parts for upgrade kits with next gen chips playing into the decision as well).

This should be factored into a decision model as well to get the best idea of what the sort of "lease" cost and period to ROI would be for different mining units, especially as the technology only continues to progress as it has.  Look up comparable fundamentals and pricing structures/strategies for Intel (INTC) and AMD (AMD) and map them to analogous demand market segments in the mining hardware space (e.g. industrial vs. individual-investor vs. hobbyist vs. curious newb) - you'll see a similar distribution of both primary and resale market share to what's going on in the Bitcoin world right now...at the end of the day, even things like brand names might command higher prices/preservation of capital value in equipment.


On a somewhat-related topic, what does everyone make of the sporadic sharp increasing perturbations in the network T operator force?  Burn-ins on pre-market hardware?

[mrlupin]

What is the "network T operator force"?

[jjdub7]

Lol, apparently the rough, rough translation for network hashrate (its an old translation joke from some tech salon announcements over the summer).  I think its literally saying the equivalent in Chinese of "network terahashes produced by the miners"

[Bonam]

I think you're missing the point of the graph. It's purpose is to illustrate, given certain parameters, what the final steady state network hashrate could end up being. Barring large changes in price or available technology, eventually the network will get to the point that even a large operation in a location with extremely cheap electricity will not be able to break even in X days even without difficulty changing.

The assumption is fundamentally invalid because there will always be changes in "available technology". Even once bitcoin mining ASICs reach the cutting edge in terms of modern process size, algorithm optimization, computational efficiency, etc, major companies in larger markets will continue to push the bounds of technology forward and ASIC manufacturers will benefit from these new technologies, releasing better and better products. The cost of a given amount of computing power is cut in half on about a yearly basis. So long as advanced human civilization exists, technology will continue to progress and there will never be a "steady state". And, so long as bitcoin remains valuable enough to warrant anyone to bother mining it with custom equipment, any miner bought at a given point in time will provide less hash rate than a miner bought a year later for the same amount of money. And that makes the graph not applicable in any real situation.

Remember where the discussion originated. I made the claim that even today, many miners fail to ever break even or just barely break even. The graph was posted as a rebuttal to that, to show that in fact capital expenditures are easy to recover and that we are far from the point where that would no longer be the case. But it only shows that in a constant difficulty environment. In a rising difficulty environment, say 10% per period, my statement is closer to the truth, as can be quickly checked with any of the conventional online mining profit calculators. And, said rising difficulty environment is the real world in which miners are trying to make back their investment and profit.

[asdfghj]

Miners improve, hence 5 lines each depicting a different J/GH.  Difficulty varies along the y-axis and is not constant in the plot.

Thanks for posting the plot, it is a nice depiction.

[electerium]

xhomer-


you are wrong.


That graph is actually a simulation of network hash rate per given variables. It could be used to calculate other things, such as profitability, but that's not its intended use and if you wanted to calculate miner profitability you wouldnt use that simulation- its redundant.

[jdany]

Is there a current AM shareholder table?

I know someone was keeping a private ledger, but, is there an official one?

The one I recall seeing had addresses, share amount and forum name (if known)

[hdbuck]

Is there a current AM shareholder table?
I know someone was keeping a private ledger, but, is there an official one?
The one I recall seeing had addresses, share amount and forum name (if known)

https://docs.google.com/spreadsheet/ccc?key=0AnxKtjWqWJGodC1fR3JDSFBpZ2tGbjBqdzJoeFlpOXc&usp=sharing#gid=0

edit: not sure where it came from tho. i just saved it a while ago.

[jdany]

Is there a current AM shareholder table?
I know someone was keeping a private ledger, but, is there an official one?
The one I recall seeing had addresses, share amount and forum name (if known)

https://docs.google.com/spreadsheet/ccc?key=0AnxKtjWqWJGodC1fR3JDSFBpZ2tGbjBqdzJoeFlpOXc&usp=sharing#gid=0

edit: not sure where it came from tho. i just saved it a while ago.

Thanks buck!
Based on the dividend, it looks like March 5th of 2014.
It would be interesting to see how shares have consolidated from 1343 direct shareholders to...

[MrTeal]

I think you're missing the point of the graph. It's purpose is to illustrate, given certain parameters, what the final steady state network hashrate could end up being. Barring large changes in price or available technology, eventually the network will get to the point that even a large operation in a location with extremely cheap electricity will not be able to break even in X days even without difficulty changing.

The assumption is fundamentally invalid because there will always be changes in "available technology". Even once bitcoin mining ASICs reach the cutting edge in terms of modern process size, algorithm optimization, computational efficiency, etc, major companies in larger markets will continue to push the bounds of technology forward and ASIC manufacturers will benefit from these new technologies, releasing better and better products. The cost of a given amount of computing power is cut in half on about a yearly basis. So long as advanced human civilization exists, technology will continue to progress and there will never be a "steady state". And, so long as bitcoin remains valuable enough to warrant anyone to bother mining it with custom equipment, any miner bought at a given point in time will provide less hash rate than a miner bought a year later for the same amount of money. And that makes the graph not applicable in any real situation.

Remember where the discussion originated. I made the claim that even today, many miners fail to ever break even or just barely break even. The graph was posted as a rebuttal to that, to show that in fact capital expenditures are easy to recover and that we are far from the point where that would no longer be the case. But it only shows that in a constant difficulty environment. In a rising difficulty environment, say 10% per period, my statement is closer to the truth, as can be quickly checked with any of the conventional online mining profit calculators. And, said rising difficulty environment is the real world in which miners are trying to make back their investment and profit.

I think there's still room to go in terms of power efficiency, but I don't know how much higher it will go. A lot of it is predicated on price. I honestly don't think there is much left to be done algorithmically to improve efficiency; the simplicity of the SHA256 means it's pretty easy to optimize. There's probably still a good bit of work to do on layout, but even that will taper out as people get designs optimized. Looking at what is currently the best implementation (Bitfury), with perfect scaling he might get ~0.2J/GH on his existing 55nm design shrunk to 28nm.
I personally think the limit of 28nm will be closer to 0.2J/GH than 0.1J/GH. You're being quite optimistic on the cost of computing power halving every year, IMO.

The biggest variable in all of this is price though. It was easy to tape out new 28nm designs and order a bunch of wafer starts when you could charge $5k for a box with <$1k in parts other than the ASIC itself. If you do get someone down to 0.2J/GH, you're already 3x better that anything else on the market at 28nm. Going from there, you're going to be looking at marginal improvements with each new spin. With margins very tight, you need to really look at what kind of income you'll generate in order to justify the cost of the new design.

I take Puppet's spreadsheet to be a simple demonstration tool, I don't think his intention is to imply it's an accurate model. I would say his assumptions are much more sound than your main premise though. If even with free electricity a miner cannot recover their capital cost, NO ONE WILL BUY MINING EQUIPMENT. Sorry for the caps, but this keeps on coming up and it's a fundamental logic flaw. With free electricity your costs of running equipment are negligible. To say that the capital costs will never be recovered assumes that income approaches 0 as the network hashrate approaches infinity.

Answer one simple question, with the assumption that continued increases in difficulty will mean that even people with free electricity can't recovery their capital costs in equipment, who do you think will be buying the ~250PH/s worth of mining hardware necessary to even double the current hashrate?

[rudi]

I don't think 28nm is the node size where we're going to hit Moore's law. 20nm chips are already being produced and TSMC will mass-produce 16nm chips in Q1 2015 (for Apple at first). I suspect the targeted 0.1J/GH are to be achieved with a smaller node size than 28nm.

[MrTeal]

I don't think 28nm is the node size where we're going to hit Moore's law. 20nm chips are already being produced and TSMC will mass-produce 16nm chips in Q1 2015 (for Apple at first). I suspect the targeted 0.1J/GH are to be achieved with a smaller node size than 28nm.

A (bad) 20nm chip was produced, but it was also funded at a time when
1. 6 month preorders were still SOP
2. Margins were extremely high
3. Mining income was still an order of magnitude higher than operational cost

Even the likes of nVidia are complaining that the economics of moving to 20nm don't make sense and that the performance gains aren't there, which is part of the reason even the new top of the line GTX980 is on the same mature 28nm node as the GPUs that debuted in Jan 2012.

Even then, Bitcoin mining is so embarrassingly parallel that it makes even a GPU look inherently serial. Unlike a GPU where two 1152-core GTX760s in SLI will perform noticeably worse than a single 2304-core GTX780, Bitcoin mining scales perfectly. There is much less pressure to move down a node as new hardware doesn't really give any new capability outside of possible efficiency gains.

Similarly to the question above, by summer 2015 if the price doesn't move Bitcoin mining hardware will be sold essentially at cost as the market saturates. Given that per gigahash 16nm isn't likely to be any cheaper than 28nm, who is going to fund the development and production of a 16nm ASIC when they would have to hope to sell a couple exahash/s worth of hardware just to recover their initial investment?

[jimmothy]

I think you're missing the point of the graph. It's purpose is to illustrate, given certain parameters, what the final steady state network hashrate could end up being. Barring large changes in price or available technology, eventually the network will get to the point that even a large operation in a location with extremely cheap electricity will not be able to break even in X days even without difficulty changing.

The assumption is fundamentally invalid because there will always be changes in "available technology". Even once bitcoin mining ASICs reach the cutting edge in terms of modern process size, algorithm optimization, computational efficiency, etc, major companies in larger markets will continue to push the bounds of technology forward and ASIC manufacturers will benefit from these new technologies, releasing better and better products. The cost of a given amount of computing power is cut in half on about a yearly basis. So long as advanced human civilization exists, technology will continue to progress and there will never be a "steady state". And, so long as bitcoin remains valuable enough to warrant anyone to bother mining it with custom equipment, any miner bought at a given point in time will provide less hash rate than a miner bought a year later for the same amount of money. And that makes the graph not applicable in any real situation.

Remember where the discussion originated. I made the claim that even today, many miners fail to ever break even or just barely break even. The graph was posted as a rebuttal to that, to show that in fact capital expenditures are easy to recover and that we are far from the point where that would no longer be the case. But it only shows that in a constant difficulty environment. In a rising difficulty environment, say 10% per period, my statement is closer to the truth, as can be quickly checked with any of the conventional online mining profit calculators. And, said rising difficulty environment is the real world in which miners are trying to make back their investment and profit.

I think there's still room to go in terms of power efficiency, but I don't know how much higher it will go. A lot of it is predicated on price. I honestly don't think there is much left to be done algorithmically to improve efficiency; the simplicity of the SHA256 means it's pretty easy to optimize. There's probably still a good bit of work to do on layout, but even that will taper out as people get designs optimized. Looking at what is currently the best implementation (Bitfury), with perfect scaling he might get ~0.2J/GH on his existing 55nm design shrunk to 28nm.
I personally think the limit of 28nm will be closer to 0.2J/GH than 0.1J/GH. You're being quite optimistic on the cost of computing power halving every year, IMO.

The biggest variable in all of this is price though. It was easy to tape out new 28nm designs and order a bunch of wafer starts when you could charge $5k for a box with <$1k in parts other than the ASIC itself. If you do get someone down to 0.2J/GH, you're already 3x better that anything else on the market at 28nm. Going from there, you're going to be looking at marginal improvements with each new spin. With margins very tight, you need to really look at what kind of income you'll generate in order to justify the cost of the new design.

I take Puppet's spreadsheet to be a simple demonstration tool, I don't think his intention is to imply it's an accurate model. I would say his assumptions are much more sound than your main premise though. If even with free electricity a miner cannot recover their capital cost, NO ONE WILL BUY MINING EQUIPMENT. Sorry for the caps, but this keeps on coming up and it's a fundamental logic flaw. With free electricity your costs of running equipment are negligible. To say that the capital costs will never be recovered assumes that income approaches 0 as the network hashrate approaches infinity.

Answer one simple question, with the assumption that continued increases in difficulty will mean that even people with free electricity can't recovery their capital costs in equipment, who do you think will be buying the ~250PH/s worth of mining hardware necessary to even double the current hashrate?

Puppets tool is not meant to predict when/howfast the difficulty will increase. It's only meant to demonstrate how the parameters affect the difficulty and give you a rough idea of what you would expect if we reached equilibrium and technology stopped advancing.

Of course technology will still advance but it won't be at the current 10%/2 weeks rate. I'd guess ~1%/diff increases if it follows Moore's law.

If even with free electricity a miner cannot recover their capital cost, NO ONE WILL BUY MINING EQUIPMENT.

Exactly this. Large scale mining operations do not have access to free electricity and need to be earning massive profits if they want to continue spending millions on NRE.

I don't understand how people can say that nobody can make a profit mining because there is some entity out there that will continue adding hashing power far beyond the point everyone is mining at a loss.

[Puppet]

Moore's law states that transistor density doubles every 18 months. Something which no longer is true, but even it where, for bitcoin transistor density is almost irrelevant at this point; so what if you can cram twice as many cores in a chip, if it consume twice as much power you havent really gained much. You reduced the silicon cost per GH, but thats arguably one of the least important metrics right now.  Moore's law says nothing about compute efficiency, and I think its fair to say the industry is heading in to a brick wall in that regard too.

Thats not to say I dont expect improvements over where we are now, there is probably still a lot to be had from improving power delivery and riding down the schmoo plot from high voltage/highest performance per mm² to low voltage/highest power efficiency, but thats pretty much going to be a one time deal.