Total Hashrate Forecast Q1, Q2 2014 (Community work)

[de_ixie]

Thank you everybody for your input. I will consolidate the data today and make a first draft. We can continue tweaking the numbers as time progresses.

[de_ixie]

Updated

[Entropy-uc]

Something to consider, if the total hash is 10 phs, and it doubles, effectively doubling difficulty, what % of the first 10 phs gets turned off?  Once the old equipment breaks even, it's house money, and I suspect only breakage or electricity>btc will be the only reason to decommission.  So does it all stay on for 2014? Or does much of it become too expensive?

At today's BTC prices you won't see anything getting turned off until we are in the 75 Ph/s region. Even then, the gear that is no longer cost effective to operate will be a small part of the total rate, so the result won't be noticeable.

A larger factor in network growth will be the time to break even on the hardware investment.  Puppet is very optimistic about price reductions, with his lowest hardware cost being less than 1/2 of the cheapest pre-order prices.  On top of that, deployments at any scale have substantial facility costs for racks, network gear, cooling and power.  These can easily approach $1000 / Th/s.

Without more massive bitcoin appreciation, the risk reward for buying hardware will look pretty ugly by the end of Q1 2014.  Paying 10s of thousands of dollars, 4 months in advance is tempting when you earn your money back in 2 weeks.  When it will take 6 months of hashing with stable difficulty and exchange rates, the smart money is going to step back.

[Puppet]

Puppet is very optimistic about price reductions, with his lowest hardware cost being less than 1/2 of the cheapest pre-order prices.  

Prices will drop that low and much lower. Lets take BFL monarch as a baseline for a second. I've  linked this table before, that shows the BOM of highend video cards:

source: Mercury Research

If you eliminate the components you dont need on a miner (mostly GPU and VRAM) you end up with $47 for a 6970. A card which has comparable power draw and size as a Monarch (but which also includes a ton or irrelevant components and connectors like crossfire, DVI, HDMI, audio, PCIe 16x, ..  and a far more complicated 14+ layer PCB to accommodate the ultra high clockspeeds and bandwidth of the VRAM).

Which leaves one to estimate the cost of the ASICs. With a per 28nm wafer price of $4000 (50% above high volume prices), a 400GH hashfast golden nonce chip would cost somewhere on the order of $30 after packaging and yield. I assume BFL and most other vendors will be in the same ballpark, order of magnitude of $100 per TH for the chips.

Combine both, and add some minimum operating margin and you will see that $1000 per TH provides these vendors with gross margins that would make intel and nvidia green with envy. In high volumes, a 600GH Monarch probably costs BFL barely over $100 in marginal production cost.

[Entropy-uc]

Puppet is very optimistic about price reductions, with his lowest hardware cost being less than 1/2 of the cheapest pre-order prices.  

Prices will drop that low and much lower. Lets take BFL monarch as a baseline for a second. I've  linked this table before, that shows the BOM of highend video cards:

source: Mercury Research

If you eliminate the components you dont need on a miner (mostly GPU and VRAM) you end up with $47 for a 6970. A card which has comparable power draw and size as a Monarch (but which also includes a ton or irrelevant components and connectors like crossfire, DVI, HDMI, audio, PCIe 16x, ..  and a far more complicated 14+ layer PCB to accommodate the ultra high clockspeeds and bandwidth of the VRAM).

Which leaves one to estimate the cost of the ASICs. With a per 28nm wafer price of $4000 (50% above high volume prices), a 400GH hashfast golden nonce chip would cost somewhere on the order of $30 after packaging and yield. I assume BFL and most other vendors will be in the same ballpark, order of magnitude of $100 per TH for the chips.

Combine both, and add some minimum operating margin and you will see that $1000 per TH provides these vendors with gross margins that would make intel and nvidia green with envy. In high volumes, a 600GH Monarch probably costs BFL barely over $100 in marginal production cost.

Your approach is good, but the numbers are bad.

AMD buys components for 100s of millions of boards. They have full time staff ordering quarters in advance and negotiating hard over every 10th of a penny (literally!).  At best mining hardware supplies will build a few 10s of thousands of systems.  At that level, the manufacturers of components won't even pick up the phone.  They will have to buy from a distributor who buys from a wholesaler with markups at every stage.  Double the component cost and you'd have a good number for the end of next year when manufacturers can negotiate based on a stable build plan.

Packaging costs for the kind of silicon being used here often higher than the silicon.  And yield can suck at both the silicon level and the packaging level, especially with the multi-chip modules being used by Hashfast and Cointerra (I have no idea what bullshit labs is doing, and will never care).  Packaged silicon costs are going to be more like $150-$200 per Th/s at best.

That just gets you to a bare board.  To get to a system level you still need cooling (high performance = high cost), a case, and power supply.  All that puts you in the range of $400 / Th/s cost at best.

Then you need to account for overhead.  Each of the professional companies have staffing on the order of 10 people, many of whom command high salaries.  Offices, health care, travel and marketing all have to be shared over the units sold.  Let's assume your typical company moves 50 Ph/s in 2014, and has $2.5M in operating expenses.  (that would be tremendous sales, and low costs).  That would add another $50 / Th/s to the system cost.

Now, how much profit is going to be acceptable?  Especially since the alternative is to keep the hardware and operate it yourself.

Prices will come down, but not as much as you think.

[Puppet]

AMD buys components for 100s of millions of boards.  

You are off by orders of magnitude. AMD sells less than 1M high end (game grade) videocards per quarter. And thats AMD, consider the bazillion OEM's that sell cards that are price competitive, yet are much lower volume, probably in same ballpark if not below some bitcoin asic manufacturers 6 months from here. Sure, AMD and nVidia will have a volume pricing advantage, but its not going to matter for this comparison, and its not going to be anywhere near a factor 2x.

Packaging costs for the kind of silicon being used here often higher than the silicon.  

Packaging is priced per ball typically. The bitcoin asics arent that high pin count, there is no way it would cost more than $10 per chip. More likely $2-5 in volume. FWIW, I calculated using $10 in packaging cost, though I think $2 is probably closer once they achieve volume's of 100K's.

As for MCM's. If it were more expensive to package 4 dies in a chip than doing 4 chips, why would they?  You might argue for watercooling that the cooling is significantly cheaper, but even KnC is using MCM's with off the shelve air cooling.

And yield can suck at both the silicon level and the packaging level

Its not going to suck for everyone all the time, which is what matters. And yield shouldnt be a problem, considering these chips have dozens of "cores" that can be disabled if defective. They should achieve very near 100% yield if you allow a few bad cores per chip. Anyway, I assumed 90% yields, which is probably low for these chips.

That just gets you to a bare board.  To get to a system level you still need cooling (high performance = high cost), a case, and power supply.

Monarch doesnt have a PSU. SHould miners include that in their ROI calculation? Yeah. WIll they? Doubtful.  Many still have piles of PSU's. And even if you dont, a PSU doesnt become worthless after one year or when mining is no longer profitable. As for case, look at KnC. Just 2 sheets of plied aluminum. Cant cost more than $10 in volume, if that. You can buy full ATX cases with fans for that much.

BTW, chip cooling is included in that BOM. In fact, its about the most expensive part of the videocard.

All that puts you in the range of $400 / Th/s cost at best.

I dont think so.

Then you need to account for overhead.  Each of the professional companies have staffing on the order of 10 people, many of whom command high salaries.  Offices, health care, travel and marketing all have to be shared over the units sold.  Let's assume your typical company moves 50 Ph/s in 2014, and has $2.5M in operating expenses.  (that would be tremendous sales, and low costs).  That would add another $50 / Th/s to the system cost.

Yes there has to be a profit margin to pay the overhead. Im calculating BOM. In most businesses a 10% margin is what it takes to operate. Put it at 20% if you want. But keep in mind, once difficulty has gone up that much that market prices approach marginal cost, the alternative for those asic vendors is shutting their doors.

Prices will come down, but not as much as you think.

Lets check back in 24 months.

[Entropy-uc]

AMD buys components for 100s of millions of boards.  

You are off by orders of magnitude. AMD sells less than 1M high end (game grade) videocards per quarter. And thats AMD, consider the bazillion OEM's that sell cards that are price competitive, yet are much lower volume, probably in same ballpark if not below some bitcoin asic manufacturers 6 months from here. Sure, AMD and nVidia will have a volume pricing advantage, but its not going to matter for this comparison, and its not going to be anywhere near a factor 2x.

Packaging costs for the kind of silicon being used here often higher than the silicon.  

Packaging is priced per ball typically. The bitcoin asics arent that high pin count, there is no way it would cost more than $10 per chip. More likely $2-5 in volume. FWIW, I calculated using $10 in packaging cost, though I think $2 is probably closer once they achieve volume's of 100K's.

As for MCM's. If it were more expensive to package 4 dies in a chip than doing 4 chips, why would they?  You might argue for watercooling that the cooling is significantly cheaper, but even KnC is using MCM's with off the shelve air cooling.

And yield can suck at both the silicon level and the packaging level

Its not going to suck for everyone all the time, which is what matters. And yield shouldnt be a problem, considering these chips have dozens of "cores" that can be disabled if defective. They should achieve very near 100% yield if you allow a few bad cores per chip. Anyway, I assumed 90% yields, which is probably low for these chips.

That just gets you to a bare board.  To get to a system level you still need cooling (high performance = high cost), a case, and power supply.

Monarch doesnt have a PSU. SHould miners include that in their ROI calculation? Yeah. WIll they? Doubtful.  Many still have piles of PSU's. And even if you dont, a PSU doesnt become worthless after one year or when mining is no longer profitable. As for case, look at KnC. Just 2 sheets of plied aluminum. Cant cost more than $10 in volume, if that. You can buy full ATX cases with fans for that much.

BTW, chip cooling is included in that BOM. In fact, its about the most expensive part of the videocard.

All that puts you in the range of $400 / Th/s cost at best.

I dont think so.

Then you need to account for overhead.  Each of the professional companies have staffing on the order of 10 people, many of whom command high salaries.  Offices, health care, travel and marketing all have to be shared over the units sold.  Let's assume your typical company moves 50 Ph/s in 2014, and has $2.5M in operating expenses.  (that would be tremendous sales, and low costs).  That would add another $50 / Th/s to the system cost.

Yes there has to be a profit margin to pay the overhead. Im calculating BOM. In most businesses a 10% margin is what it takes to operate. Put it at 20% if you want. But keep in mind, once difficulty has gone up that much that market prices approach marginal cost, the alternative for those asic vendors is shutting their doors.

Prices will come down, but not as much as you think.

Lets check back in 24 months.

In 24 months it won't matter.  You were talking about the next year.  In 24 months if bitcoin survives, there will be 14 nm devices on the market and pricing will be totally different.

Your packaging numbers are reasonable for 5W chips.  For the 500W, GHz devices we are talking about now packaging is going to be more on the order of $25++.  Getting clean signal and power into these devices is difficult at high clock speeds.

Good point about the size of the card market. I overlooked AMD acting as a foundry in that space. It still leaves miner hardware manufacturers orders of magnitude too small for the pricing you anticipate.

I don't think any hardware manufacturer will operate with 10% margins.  This is not a commodity market.  It's an extremely specialized niche with a volatile underlying valuation and huge regulatory risks.  Investors will be looking for greater than 50% margins AND offloading of risks to customers through pre-payments.

Finally, there aren't enough hobby miners to support the suppliers currently in the market.  So, yes, all in costs per Th/s will absolutely apply to purchase decisions going forward.

[Puppet]

Keep in mind Im estimating the "end game". If you just look at the PH range Im getting, bitcoin asics will be a big market, arguably as big as many high end gaming card manufacturers. 500 PH = over 1 million "Golden nonce" chips or 20 million Bitmine A1s. Thats high enough volume to get you the discounts that matter.

About the packaging.. 500-700 MHz isnt that much these days. ARM mobile chips do a lot more than that, and they still cost only ~$5 in volume. Die + packaging + odm margin. Sure, they are low power, but they also have far more complicated IO requirements. A bitcoin asic needs next to no IO. Look at bitmine's chip design, just a handful of (large) pads.

As for the OEM margins; what choice will they have? They can keep prices as high as they want, but if these chips are no longer operationally profitable to mine with, they wont sell any. If they are only barely operationally profitable, miners wont be interested in investing something that wont ROI in 10 years. This is not like GPU or CPU market. bitcoin OEMs will have to keep dropping prices, following difficulty (divided by bitcoin price)  or at some point, close their doors. So their BOM is a good indication of how low prices will eventually fall. Its not because these vendors currently can charge phenomenal prices that it has anything to do with cost. Its just because these chips are for now, phenomenally "profitable". Obviously that wont last.

[smracer]

Josh from BFL just said this yesterday:

_______________________________________________________________________________ ________________________________
No, there's not 100,000,000 Monarchs in the queue... less than 10k I believe. It will not take 100 years to fill.

If you order your Monarch today, I would expect it to be shipped sometime toward the end of March or beginning of April, possibly middle of April. It all depends on how long it takes to get the second batch of chips out of the foundry and ready to go, as well as any problem encountered with chip turn up, if any (hopefully none).

Nothing I have written here is any sort of guarantee on dates and merely speculation based on currently available information.
_______________________________________________________________________________ ________________________________

10K max Monarch's ordered.  That is 6PH.

[Puppet]

And you believe him? Just the fact that he says "I believe" is enough to dismiss it. As if he wouldnt know how many they sold...
Its also kinda hard to rhyme this statement with another statement he made a while ago, claiming they would be able to assemble 1K Monarchs per day. If they really sold only 10K, thats ~10 days. Why would that take until April by his own always-incorrect forecast-not-a-promise?

[de_ixie]

Small update from my side:

- Starting point of 11Ph seems ok

- Hashfast failed to provide their initial batch on time; Lawsuit upcoming; Lets see whether they survive the upcoming months; My feeling is they will somehow scam out of the mess they've caused and survive - but at the cost of a heavily damaged reputation; They will provide less than we estimated, and later....; For now - until clarification of situation - lets keep FC like it is; We have potential to decrease our FC

- Some unknown source keeps adding massive hashpower despite no major vendor is shipping; We have to observe that trend; I am not sure what is behind that. Maybe just the sum of small various deployments (would not be a prob)? Or maybe a hidden project (would be a major prob)? My feeling is we have to increase our "Others" position soon

- We have a random statement of Josh from BFL claiming Monarch preorder count below 10k (equals ~6 PH). My feeling is we might soon have potential to decrease our BFL estimate

Next review around end of January; Earlier in case something big is happening. By end of this month, we will have a lot more information about some crucial milestones; Asicminer (timely tapeout of their Chip?), Bitmine (can they show a prototype?), Hashfast (will they bring their shit together?), blackarrow (nothing neg. heard...), Cointerra (do they deliver/ show a working product), KnC (any Milestones?) etc...

Stay tuned

[FeedbackLoop]

Thanks for the excellent continuation of this thread de_ixie!


Here's a tiny update: Cointerra now sold out April and taking May orders:

http://cointerra.com/shop/

[de_ixie]

Thanks for the excellent continuation of this thread de_ixie!


Here's a tiny update: Cointerra now sold out April and taking May orders:

http://cointerra.com/shop/

Thx - updated

[Cablez]

One of the dark farms has a name at least.  Called '21e6' from here of all places (3/4 of the way down):  http://www.businessweek.com/articles/2014-01-09/bitcoin-mining-chips-gear-computing-groups-competition-heats-up


Big backers so could be huge.

[Bismarckbkk]

One of the dark farms has a name at least.  Called '21e6' from here of all places (3/4 of the way down):  http://www.businessweek.com/articles/2014-01-09/bitcoin-mining-chips-gear-computing-groups-competition-heats-up


Big backers so could be huge.

 "Occasionally, he says, he stuffs the air holes of his machines with paper to bring up the temperature" haha

[ImI]

One of the dark farms has a name at least.  Called '21e6' from here of all places (3/4 of the way down):  http://www.businessweek.com/articles/2014-01-09/bitcoin-mining-chips-gear-computing-groups-competition-heats-up


Big backers so could be huge.

 "Occasionally, he says, he stuffs the air holes of his machines with paper to bring up the temperature" haha

i am still not convinced that it brings up the hashrate

[Bismarckbkk]

Of course it doesnt, its completely ridicolous. Worst case it can damage the hw.

[brontosaurus]

Puppet is very optimistic about price reductions, with his lowest hardware cost being less than 1/2 of the cheapest pre-order prices.  

Prices will drop that low and much lower. Lets take BFL monarch as a baseline for a second. I've  linked this table before, that shows the BOM of highend video cards:

source: Mercury Research

Which leaves one to estimate the cost of the ASICs. With a per 28nm wafer price of $4000 (50% above high volume prices), a 400GH hashfast golden nonce chip would cost somewhere on the order of $30 after packaging and yield. I assume BFL and most other vendors will be in the same ballpark, order of magnitude of $100 per TH for the chips.

I like this thread, it's good work and an excellent community resource. I hate to be a party pooper, but your wafer prices are way too low by a large margin. Qualcomm might get an untested 28nm wafer for $3500, but no rig vendor is getting their 28nm wafers any less than about $20k a pop - wafer fabs can sell more 28nm than they can make just now, the Bitcoin asic business is simply background noise to them.

[Puppet]

I like this thread, it's good work and an excellent community resource. I hate to be a party pooper, but your wafer prices are way too low by a large margin. Qualcomm might get an untested 28nm wafer for $3500, but no rig vendor is getting their 28nm wafers any less than about $20k a pop - wafer fabs can sell more 28nm than they can make just now, the Bitcoin asic business is simply background noise to them.

Nonsense. First of all, fabs are idling.

Taiwan Semiconductor Manufacturing Company (TSMC) has seen its 28nm process utilization rate fall to 65-70% recently, due to a cutback of customer orders, according to ind
http://www.digitimes.com/news/a20131209PD213.html

30-35% of their capacity is unused. Hardly "they can sell more then they can make" is it.

Your price estimates are even further off. Volume prices is below $3000 nowadays.  Low volume prices through intermediaries may be much higher, but thats irrelevant when estimating the "end game". To get to anywhere near the hashrates I predict, you are talking about significant volumes of wafers anyway.

[brontosaurus]

I like this thread, it's good work and an excellent community resource. I hate to be a party pooper, but your wafer prices are way too low by a large margin. Qualcomm might get an untested 28nm wafer for $3500, but no rig vendor is getting their 28nm wafers any less than about $20k a pop - wafer fabs can sell more 28nm than they can make just now, the Bitcoin asic business is simply background noise to them.

Nonsense. First of all, fabs are idling.

Taiwan Semiconductor Manufacturing Company (TSMC) has seen its 28nm process utilization rate fall to 65-70% recently, due to a cutback of customer orders, according to ind
http://www.digitimes.com/news/a20131209PD213.html

30-35% of their capacity is unused. Hardly "they can sell more then they can make" is it.

Your price estimates are even further off. Volume prices is below $3000 nowadays.  Low volume prices through intermediaries may be much higher, but thats irrelevant when estimating the "end game". To get to anywhere near the hashrates I predict, you are talking about significant volumes of wafers anyway.

When did you llast see a quotation or invoice from TSMC or Global Foundiries for 28nm?