Bitcoin: On our way to a Petahash

[bcp19]

Where did you get the $5k price on 28nm wafers?  

Experience and industry contacts

edit, found a public source for you:
TSMC decided to increase prices for 28nm chips "by somewhere between 15% - 25%. Given that a wafer processed using latest process technologies can cost $4000 - $5000, ..
http://www.xbitlabs.com/news/other/display/20110912192619_TSMC_Reportedly_Hikes_Pricing_on_28nm_Wafers_Due_to_Increased_Demand.html

Please note the date on that, 2011, when 28nm was brand new, very low volume and very few fabs offered it. 28nm is now mainstream, offered in volume by all major fabs. So If anything,  $4000 today would be a high estimate, but Im okay using it since Cointerra is not an AMD or nViida and pretty low volume for a company like TSMC.

Does that 5K include the design cost of the masks?  

Absolutely not. Masksets, especially for 28nm, are orders of magnitude more expensive, but its a one time, sunk cost. Once you have a maskset, to determine production volume all that matters is the variable production cost vs market value. Sunk costs will determine whether or not you end up making a profit, but not whether or not you should produce more batches. Its like when buying a miner, once you bought it, purchase price (sunk cost) no longer matters to decide to run it or not, you will run it until it earns you less than it costs in electricity, but the purchase price will largely determine if you end up making a profit or not.

What about the bumping and binning processes?  The boards? Etc?

"Bumping and binning" is part of testing and packaging. Its not included in the $5K/wafer price, but it is included in a 'a few dollar per chip", since the cost is typically closer related to the number of die's than wafers.
If you want a public reference for that too, this is all I can find quickly:
http://electroiq.com/blog/articles/2001/03/technology-comparisons-and-the-economics-of-flip-chip-packaging/

(scroll down to the last charts as the others are unreadable, and you may want to note the numbers they picked for average die cost too: $5. Not $50 much less $500).
Please note that article is more than 10 years old. Gives you an idea what BGA packaging costed back then, when it was state of the art technology. The numbers have dropped since, but are still within the same ballpark of a few dollar per chip.

For PCBs, you can get an idea from the prices of infinitely more complicated PC motherboards (that contain dozens of other asics) and still sell for $20 or so in bulk, or from a raspberry Pi that retails for $40 including the  SoC, RAM, IO etc.  
PCB cost doesnt amount to much in the big picture.

Sorry, I just don't see it.

You can lead a horse to water...

Hmm, let's take a look at avalon with your numbers.  We know Avalon produced 3 batches of product and took orders for 1,000,000 chips.  I can't seem to get a solid number, but 400MH per chip seems likely.  So, we have 1500 units operating at ~65GH  which gives up ~164 chips per unit.  All in all that comes to 1,246,000 chips.  At 1188 chips per wafer, that cones to ~1050 wafers.  Those 1050 wafers would run ~$4.2M at $4k each or ~$5.2M with the masks.  $4.17 per chip or $684 per unit.   $1299/75btc/75btc = $389,700/45kbtc/45kbtc = $911,700/$3,476,250.  $4,777,650 income from the 3 batches.  78000 BTC = ~$10,920,000 for the 1,000,000 chips.  $15,697,650... I can see Avalon taking a 3x markup overall, even though it was mostly B3 that got screwed.

BFL went with 65NM and seems to have ~4GH per chip.  They must have a bigger chip size or did a better job with the traces, as halving the technology size should only give a 4x increase in traces.  Going to 28NM would roughly involve another 4x increase in traces.  6.4GHz off of a ~4.65x4.65mm chip (16.13mm²).  A single 500GH chip would require nearly 78x as many traces, or 35.5mmx35.5mm which wuld rival the size of an Intel processor which is 37.5x37.5, in which case using a 300MM wafer you'd get 48 chips.  At $4k per wafer, that's $83.33 per chip.

[1714917746]

1714917746

[1714917746]

1714917746

[1714917746]

1714917746

[Puppet]

 A single 500GH chip would require nearly 78x as many traces, or 35.5mmx35.5mm which wuld rival the size of an Intel processor which is 37.5x37.5, in which case using a 300MM wafer you'd get 48 chips.  At $4k per wafer, that's $83.33 per chip.

Arhm. Nope.
 35mmx35mm would dwarf any CPU intel has ever produced, including the highest end Itaniums.  You do realize that would be a 1250mm2 chip right? Thats 10x the typical cpu size. I think you are confusing die size with package size.

Anyway, hasfast is on record with their die size:
https://bitcointalk.org/index.php?topic=262052.msg2975190#msg2975190

18x18 = 324 mm^2 for 400GH.
Thats a big chip, perhaps bigger than I expected, but interestingly it matches exactly my high end estimate (hadnt seen that post when I made my calculations).
That still yields 177 candidates or $22 per chip candidate at $4K/wafer.

Whether thats actually good or bad compared to cointerra, BFL and the others remains to be seen, but its the first, and a useful datapoint that requires no guessing.

Ill be fair, Hashfast is not likely able to sell those chips packaged and functional with PCB and everything at $30 with a profit, but even if you assume hashfast will be among the best in GH/mm2 and you adjust my numbers to $50 per chip or so, it hardly changes the big picture. And if you fast forward 12 months from here, $2500-$3000 is a more likely wafer price for large volumes.


edit: one more point. That post I linked to is actually utter nonsense when it comes to KNC's chip. You cant deduce die size by looking at a (rendering) of a heatspreader. Take any Core or Pentium chip, and remove the heatspreader to check the naked die size. You will find its anywhere from 50% to 15% the size. That a "VP of engineering" would spout such nonsense doesnt make me feel very good about hashfast. ALl you can say with confidence is that the die cant be bigger than the heatspreader (duh), but I would be very surprised if KNC were indeed making a 1000+mm2 behemoth.  Not even AMD and nVidia dare to go that big with their top end GPUs.

[bcp19]

Ah, indeed, I was looking at packaging size. 

Does this mean Avalon's chips were very very inefficient?  Or is there a scaling I am missing?  28nm tech is basically 1/4th x 1/4 or 1/16th the area for the same traces.  I had been looking at the back of GTX480 I removed from a system a while back, and chip looked to be almost 1 1/2" square, so the 37.5 seemed feasible .  On the CPUs, is the package size so large to accomodate all of the contacts?

[Puppet]

Does this mean Avalon's chips were very very inefficient? 

The process nodes are so far apart, its really an apples to oranges comparison. But lets see what we get.

Avalon 16 mm2 ~300MH @110nm
~19MH/mm2

Hashfast is 324 mm^2 for 400GH @28nm
~1234MH/mm2

From 110nm to 28nm youd expect a 15x increase in transistor density (110/28)^2.
Hashfast actually promises a 64x speedup per mm2 compared to avalon, or ~4x more than expected from shrinking alone.

Most, if not all  of that, can be explained by increased clock speed  (smaller node=faster switching frequency).
Also, hashfast has yet to prove its claimed performance, so lets wait and see.

But even so, it should come as no surprise that the first ever bitcoin asic to hit the market wouldnt also be the most efficient one possible.
I also doubt hashfast will be very close to that, but we will have wait for the others to compare.

  I had been looking at the back of GTX480 I removed from a system a while back, and chip looked to be almost 1 1/2" square, so the 37.5 seemed feasible .  On the CPUs, is the package size so large to accomodate all of the contacts?

Yes, its mostly to accommodate all the the IO and power pads.

As for your GTX480:

GF100 - 529.17mm2 for every GeForce GTX 465, 470, 480, 485[?]
Currently, GF100 is the largest monolithic chip in the world

http://www.brightsideofnews.com/news/2010/8/9/nvidia-fermi-geforce-die-sizes-exposed.aspx

Thats a big sucker. The above quote is actually incorrect though, image sensors can be much larger than that, all the way to wafer scale sensors for stuff like mamography.
But for "asics" (cpu, gpus and the like) anything over 300mm2 is generally considered big and over 500mm2 huge.

[bcp19]

Does this mean Avalon's chips were very very inefficient? 

The process nodes are so far apart, its really an apples to oranges comparison. But lets see what we get.

Avalon 16 mm2 ~300MH @110nm
~19MH/mm2

Hashfast is 324 mm^2 for 400GH @28nm
~1234MH/mm2

From 110nm to 28nm youd expect a 15x increase in transistor density (110/28)^2.
Hashfast actually promises a 64x speedup per mm2 compared to avalon, or ~4x more than expected from shrinking alone.

Most, if not all  of that, can be explained by increased clock speed  (smaller node=faster switching frequency).
Also, hashfast has yet to prove its claimed performance, so lets wait and see.

But even so, it should come as no surprise that the first ever bitcoin asic to hit the market wouldnt also be the most efficient one possible.
I also doubt hashfast will be very close to that, but we will have wait for the others to compare.

  I had been looking at the back of GTX480 I removed from a system a while back, and chip looked to be almost 1 1/2" square, so the 37.5 seemed feasible .  On the CPUs, is the package size so large to accomodate all of the contacts?

Yes, its mostly to accommodate all the the IO and power pads.

As for your GTX480:

GF100 - 529.17mm2 for every GeForce GTX 465, 470, 480, 485[?]
Currently, GF100 is the largest monolithic chip in the world

http://www.brightsideofnews.com/news/2010/8/9/nvidia-fermi-geforce-die-sizes-exposed.aspx

Thats a big sucker. The above quote is actually incorrect though, image sensors can be much larger than that, all the way to wafer scale sensors for stuff like mamography.
But for "asics" (cpu, gpus and the like) anything over 300mm2 is generally considered big and over 500mm2 huge.

Thank you.  You've pointed out a lot of information that I had previously been unable to find in my searches.  I have a clearer picture of what you are talking about now, but I do find one small flaw.  Look at what has happened to the market since the introduction of ASICs.  Sure, the first few would make money fast, but others would follow and to keep market share you'd have to continue spending more and more.  Soon you'd be up to 30-60 days RoI and climbing.  Why mine with something when you can get more by selling it?  The other thing to consider... what would happen to the price of BTC if 2-3 companies were mining the majority of it?  Businesses need capital to function.  They'd have to sell and sell and sell.  Who'd want to spend hundreds of thousands of dollars daily to keep BTC prices high while the big businesses profit?

[Puppet]

Thank you.  You've pointed out a lot of information that I had previously been unable to find in my searches.  I have a clearer picture of what you are talking about now, but I do find one small flaw.  

Impossible. My reasoning and posts are widely known to be flawless

Look at what has happened to the market since the introduction of ASICs.  Sure, the first few would make money fast, but others would follow and to keep market share you'd have to continue spending more and more.  Soon you'd be up to 30-60 days RoI and climbing.  Why mine with something when you can get more by selling it?

 

I dont understand your point, much less how it would contradict mine.

The other thing to consider... what would happen to the price of BTC if 2-3 companies were mining the majority of it?  Businesses need capital to function.  They'd have to sell and sell and sell.  Who'd want to spend hundreds of thousands of dollars daily to keep BTC prices high while the big businesses profit?

Potential 51% attack  issues aside, it doesnt matter if its one company or 10000 individuals mining. The amount of coins mined remains the same, and freshly minted coins are only a very small % of all the bitcoin trade volume anyway. Its pretty much irrelevant if miners hold or sell. The two largest dollar/BTC exchanges alone have a volume of nearly 1M BTC per month. Miners mine what, a bit over 100K BTC per month?  Not gonna make the difference.

[bcp19]

Thank you.  You've pointed out a lot of information that I had previously been unable to find in my searches.  I have a clearer picture of what you are talking about now, but I do find one small flaw.  

Impossible. My reasoning and posts are widely known to be flawless

Look at what has happened to the market since the introduction of ASICs.  Sure, the first few would make money fast, but others would follow and to keep market share you'd have to continue spending more and more.  Soon you'd be up to 30-60 days RoI and climbing.  Why mine with something when you can get more by selling it?

 

I dont understand your point, much less how it would contradict mine.

The other thing to consider... what would happen to the price of BTC if 2-3 companies were mining the majority of it?  Businesses need capital to function.  They'd have to sell and sell and sell.  Who'd want to spend hundreds of thousands of dollars daily to keep BTC prices high while the big businesses profit?

Potential 51% attack  issues aside, it doesnt matter if its one company or 10000 individuals mining. The amount of coins mined remains the same, and freshly minted coins are only a very small % of all the bitcoin trade volume anyway. Its pretty much irrelevant if miners hold or sell. The two largest dollar/BTC exchanges alone have a volume of nearly 1M BTC per month. Miners mine what, a bit over 100K BTC per month?  Not gonna make the difference.

While there may be 1M BTC traded each month, you're not looking at the bigger picture.  First, the majority of what is traded is kept on the market.  Mostly you have people either buying from Bots to get a few coins for whatever they want to use elsewhere or selling to other bots to get some cash out.  Unless the $ in equal $ out, you're going to have a price increase or decrease.  Let's say there's $10M in Mt Gox.  If Company XYZ is mining 1000BTC daily and selling them on MT Gox daily, at current prices that'd be around $138,000.  In 72 days that $10 mil is wiped out unless the price drops or people start sending $140k per day to Mt Gox to offset the withdrawls.  Which do you think is more likely, a price drop or people willing to add $140K a day to keep the BTC price stable?  It might happen 1 day or 2 days or even 3 days... but what about day 15?  day 30?  day 60?

Once the price drops, adding new hardware to the network results in less gain.  Sooner or later power will catch up and then where are you?

If you still don't see it, then imagine Bill Gates deciding to sell 1000 shares a day.  Once wall street finds out Bill is selling, who the heck is going to want to buy?

[Puppet]

I dont see what this has to do with anything or at least with my point that bitcoin mining hardware production cost is almost two orders of magnitude below current market prices, and that the inevitable conclusion of this is that difficulty will rise by a comparable amount in the time it takes these vendors to produce those amounts.

 I suggest you discuss the effects of miners selling off in the bitcoin speculation threads, but I dont see why concentrated mining power would behave differently from distributed mining when it comes to selling bitcoins, nor do I see how anything I said would lead to more concentrated mining.

[bcp19]

I dont see what this has to do with anything or at least with my point that bitcoin mining hardware production cost is almost two orders of magnitude below current market prices, and that the inevitable conclusion of this is that difficulty will rise by a comparable amount in the time it takes these vendors to produce those amounts.

 I suggest you discuss the effects of miners selling off in the bitcoin speculation threads, but I dont see why concentrated mining power would behave differently from distributed mining when it comes to selling bitcoins, nor do I see how anything I said would lead to more concentrated mining.

You said the companies would either sell or produce and mine themselves.  I can see the former, but I dispute the latter for the reason stated above.  The majority of people buying mining equipment are not into the get-rich-quick scene.  If they were, you'd be seeing a lot more refunds from all companies as their chance at RoI vanishes.

People kept buying GPUs way back when while the RoIers were screaming it wasn't worthwhile.  Poeple are buying ASICs today while the RoIers scream.  Bitcoin is not made up solely of the RoIers, it's made up of a lot of individuals looking at it long term.  Businesses cannot afford the long term outlook, they need working capital on a regular basis.  That is why half of your conclusion fails, the buinesses cannot afford to mine for themselves or they will destroy the market.

Maybe you need another viewpoint to understand.  Take Magic the Gathering trading cards.  If I buy 100 boxes of the newest edition and sort them out, I can then sell the rare cards and make a profit.  But, if WOTC decided to just print everything out and sell the individual cards people would never pay as much.  They would be stuck with thousands upon thousands of cards that no one would want to buy and the ones that people did want would never command the same prices as they would if individuals were opening the packs.  The ASIC companies are selling the equilavent of trading card packs.  When we mine the coins, they have meaning, just like if we found that one particular rare card in a pack.  If the ASIC companies mined all the coins and then offered to sell them to us, they'd not mean as much and the value would drop.

We'll never hit 1 ezahash without something changing.  The current price of BTC would never sustain it.  Using your $50 for 500GH, it'd cost $99.9 million to get from where we are today to 1 exahash.  We're minting ~600k a day at current levels and that's not including power costs.  1 exahash running at .7 watts per GH would consume 700 megawatts an hour, 16.8 gigawatts per day.  At a mere $.05 per kilowatt you're looking at $840,000 in power costs per day.  -$240k per day... not going to happen.

[Puppet]

We'll never hit 1 ezahash without something changing.  The current price of BTC would never sustain it.  Using your $50 for 500GH, it'd cost $99.9 million to get from where we are today to 1 exahash.  We're minting ~600k a day at current levels and that's not including power costs.  1 exahash running at .7 watts per GH would consume 700 megawatts an hour, 16.8 gigawatts per day.  At a mere $.05 per kilowatt you're looking at $840,000 in power costs per day.  -$240k per day... not going to happen.

At $.05, no, but in Russia electricity can be had for half that price. In Kuwait its only once cent

Anyway, one exahash may not happen if Hashfast is among the most efficient design, but that remains far from certain.  Most miner gear is being tuned for performance right now, not power efficiency. But you can make your own simulations, I put up a spreadsheet

https://bitcointalk.org/index.php?topic=295270.msg3165594#msg3165594

Here is chart from it:

[bcp19]

We'll never hit 1 ezahash without something changing.  The current price of BTC would never sustain it.  Using your $50 for 500GH, it'd cost $99.9 million to get from where we are today to 1 exahash.  We're minting ~600k a day at current levels and that's not including power costs.  1 exahash running at .7 watts per GH would consume 700 megawatts an hour, 16.8 gigawatts per day.  At a mere $.05 per kilowatt you're looking at $840,000 in power costs per day.  -$240k per day... not going to happen.

At $.05, no, but in Russia electricity can be had for half that price. In Kuwait its only once cent

Anyway, one exahash may not happen if Hashfast is among the most efficient design, but that remains far from certain.  Most miner gear is being tuned for performance right now, not power efficiency. But you can make your own simulations, I put up a spreadsheet

https://bitcointalk.org/index.php?topic=295270.msg3165594#msg3165594

Here is chart from it:

I can't see all the bitcoiners moving to Russia or Kuwait

[Puppet]

I can't see all the bitcoiners moving to Russia or Kuwait

In Russia mining would remain profitable at network speeds where its unprofitable elsewhere; can you not see that russians would keep buying profitable mining rigs? Can you not see that europeans or US miners would shut down theirs rigs or sell them to a russian? The people wont move, but the mining rigs will end up where its most profitable.

[Syke]

I can't see all the bitcoiners moving to Russia or Kuwait

In Russia mining would remain profitable at network speeds where its unprofitable elsewhere; can you not see that russians would keep buying profitable mining rigs? Can you not see that europeans or US miners would shut down theirs rigs or sell them to a russian? The people wont move, but the mining rigs will end up where its most profitable.

Absolutely. In the long run, miners will locate where low cost electricity is (as long as a good Internet connection is also available).

[Puppet]

18x18 = 324 mm^2 for 400GH.
Thats a big chip, perhaps bigger than I expected,

Turns out its not such a large die after all; its an MCM, multichip module that constist of 4x 81mm² dies, which is much closer to the size I had expected.
https://bitcointalk.org/index.php?topic=262052.msg3190992#msg3190992

Interestingly, calculating the number of chip candidates with this size yields a smaller TH/wafer potential. I would have expected the opposite, smaller squares can better fill a circle,  but the default values for spacing between dies more than offset that gain.
Its only a few %, not worth redoing my math for, probably a bigger factor is the MCM packaging which will cost more than traditional asics.

[ThePanCakeKid95]

Well we are now at 3 PetaHashes 

[rampalija]

not so sure about PETAHASH, only thin is wait and see

[ThePanCakeKid95]

4 Petahashes, gosh Bitcoin's hashrate is exploding