HashFast launches sales of the Baby Jet

[MrTeal]

I would be interested to know what method the dies are attached to the heatspreader.

[amer]

I know it is the obligation of everyone in the forum to chase everyone off (moar BTC for me!) , so I know an optimistic outlook is a faux pas,

As if that would work. Many of us have been predicting this bloodshed among virtually all asic customers for years now, it didnt change a thing, and wont change a thing.  

I'm not saying it works, I'm saying that's the attitude, which I guess you are agreeing with. There are posts on even this thread about how KnC will work out, but everyone else is doomed. Meanwhile, KnC had this exact same panic cycle right before launch, complete with lawsuits and is still not clearly "not doomed".

but there is a cap on how many people want to buy ASICs and we're going to hit an inflection where hardware supply will simply exceed demand.

Doesnt everybody want to make money without working? As long as asics are at least marginally profitable, people will buy.

Of course, that's the whole point of this, no one is mining for any other reason. However, "everybody" needs to also believe that BTC are a thing and this community is not that big. We are finite and as unfathomable as it sounds and there is a point where we have gotten our hands on as many miners as we are able to deal with and the hardware manufacturers will run out of quantum leaps to sell to us and do things like "have inventory". Not to mention, miners don't just have to compete with each other, but they have to compete with "buying BTC and just holding onto it" which has the cost of 0 cents per kWh.

And when people no longer buy, prices will just drop until they buy again. The capacity is there, its going to be used one way or another. Even if miraculously miners no longer buy, these asics will just end up in large private mines.

We're agreeing. The point at which supply exceeds demands is an example of one reason to lower prices - to increase demand again.

BTW, Im sure there are loads of (ex)miners like me, sitting out the current onslaught, but who want to get in to the game again once things have settled down a bit, say next summer or so. I dont mind thin margins and long ROI times, I just dont want it to be a 100% gamble if my chosen manufacturer will be shipping one week sooner or later and how much the others will ship in the 6 weeks thereafter. Satoshi dice gives you far better odds. But that doesnt mean I wont buy an asic ever, if the opportunity presents itself, Ill buy. But thats not gonna happen anytime soon.

Also a great point. I'm saying is that there is a limit to how many miners people will buy and that is bound by common sense (like you), your available electrical capacity (home or what you're willing to host somewhere), the BTC market conditions (CURRENTLY not being a factor, but not a guarantee) and physics (28mm is the min size for BTC ASICs for the near term). 650PH predictions make no sense because of all these reasons and are based on "well, if it keeps going on forever like it has for just the last two months then this is where we'll be" and not on all the things you are bringing up. The last two months are not normal and long-term predictions based on a very unique period in BTC history is what everyone is doing and its misguided.

[Puppet]

650PH predictions make no sense because of all these reasons and are based on "well, if it keeps going on forever like it has for just the last two months then this is where we'll be" and not on all the things you are bringing up.

FWIW, I end up with very similar numbers when calculating the "end game", using todays BTC value and estimating costs based on HF's chip efficiency and estimated production costs:



The big unknown is how fast we will get there, it will probably take longer than a year, but I have no doubt we will get there if todays BTC value holds.

[Ytterbium]

unrolled cores are not necessarily crappy cores.

Agree however they can't be compared directly.   It takes 80 clock cycles for a Bitfury rolled core to complete a hash where as an unrolled core will complete a hash in one clock cycle. So clock for clock Bitfury 756 rolled cores complete about the same number of hashes are 9 unrolled cores.  This isn't to say one method is better than another.  They are just different.  Given similar die efficiency a 9 unrolled core chip and a 756 rolled core chip would be comparable size, power usage, and hashpower.

Another way to look at it is GN's 768 nonces per clock cycle would be the equivalent of Bitfury announcing they have a new processor with 61,440 rolled cores.

Any way you slice it a single GN processor is a staggering number of cores each one processing a full nonce every clock cycle.   No existing design comes close to putting as much hashpower in a small package (Cointerra probably will be similar but so far specs are unknown).

And much of that can be expressed as a GH's / Die area, that would give you a measure of efficiency - or in its simplest terms, GH's/S. per square mm of die area - at least when comparing like for like (all 28nm's for instance)

Agreed and by that metric the GN is extremely dense.   A huge amount of hashing power per mm2.

I remember reading that the Hashfast was about 6.5x6.5mm, or 42mm²   The bitfury chips are only 3.8x3.8, or 14mm², about 3 times the surface area.

According to this it only takes 65 cycles for a bitfury core to compute a hash, not 80.  

1. 756 double sha256 cores. 61+4 kernel (61 clock cycle computation 4 clock cycle load).

2. There's asynchronous 'match' signal - the only thing that core sends out. And some busses to load data.

3. wirebond. die is laid normally in cavity. i.e. it is not flip-chip and not arranged to give heat into anything else, but PCB.
It is actually not complex to dissipate 3W... Maybe even 5W with metal-core PCB and proper cooling. That's what we'll see.

Also, the theoretical speed of the bitfury chips (which is what we're comparing) was about 5Gh/s.  40nm²/28nm² = 2.25, so HF should be able to cram 2.25 times twice as many transistors in the same amount of space. And those transistors should be faster, with about 0.75 times the transistor gate capacitance.

So, if you were able to scale the bitfury design to 28nm, and increase the surface area to 42mm², you would expect the theoretical speed to equal 2.25*1.5*3* = 50.6Gh/s.  On the other hand, it would only take about 33W as well.

It's interesting that the W/Gh/s would actually be a little better, for bitfury while the surface area efficiency seems to be about 10x for HashFast.

It could be that the BF chips may have been deliberately spread out in order to make it easier to keep the chips cool, obviously the HF chips require a lot more cooling. It's actually likely that complete miners are actually cheaper to build and deploy using bitfury chips due to the reduced cooling requirements.

[aerobatic]

Ytterblum -

i remember reading the hashfast chip had 4 dies each of 9x9 mm...  thus a total of 324mm² to get a nominal 400 GH/s and a max of 540 GH/s, thus their perf per mm is between 1.23 GH/mm²  (nominal) and 1.66 GH/mm²  (max overclock and not recommended to be run at this speed, as mentioned in their latest specs)

comparing 40nm to 28nm isn't a linear comparison like you did.  its two dimensional, thus maybe you could fit four times (or at least more than two times since 40 isn't a doubling of 28) as many transistors in a similar die area.

also, don't forget that the bitfury chip is 'full custom' which means he laid out his circuit by hand - painstaking work thats not likely to be repeated very often.  all the other bitcoin chips are standard cells using someone else's cell library (apart from the vmc/amc, which is an eAsic, and thus can probably be written off in perf and power terms)

-- Jez

[DeathAndTaxes]

Ytterbium,
Thanks for the correction on rounds (65 vs 80) and for the link.  Link saved.  I assumed 80 because SHA-2 is a 80 round cipher they must do some preprocessing or optimization which makes sense and it roughly fit with likely frequency range.

As aerobatic stated the GN die efficiency is 1.23 GH/mm² (nominal), 1.66 GH/mm² (overclocked).

Saying Bitfury has a theoretical speed of 5 GH/s is kinda pointless, it was the design goal but never acheived in real world not even once under lab conditions.  Real world Bitfury is more like 1.5 GH/s (nominal), 3.0 GH/s (overclocked)*.  That gives Bitfury (@55nm) a die efficiency of 0.10 GH/mm2 (nominal), 0.21 GH/mm² (overclocked).  You stated 40nm but Bitfury is actually 55nm, and the efficiency (both GH/mm2 and J/GH) are impressive for 55nm**.  The numbers might look low to some reading but that is the power of a couple doublings.   To show you some bad efficiency BFL for example is 65nm but lets boost their stats by 40% (65^2/55^2) to put BFL and Bitfury on the same 55nm process node.  BFL's die efficiency is 0.062 GH/mm2, with a 40% boost for 55nm vs 65nm for apples to apples it is still only 0.087 GH/mm².  Now that is with the chips overvolted and driven pretty hard and hot.  Ouch BFL.

So how would a Bitfury @ 28nm compare to HashFast Golden Nonce?   Note: this shouldn't influence anyone purchase decision as we don't even know if Bitfury plans a 28nm, when it will be released, if they will hand place it, and if these theoretical gains are possible.  With that caveat, a die shrink from 55nm to 28m means 4x the transistor density (55²/28² = 4.0).  Lets be generous and say 50% higher clocks are possible (real world is probably going to be less than what raw capacitance and switching time would indicate but I am trying to err towards the upper limit)  Between clock increase and transistor density performance is likely going to be capped at 6x.  A theoretical "Bitfury28" would be ~0.60 GH/mm² (nominal), 1.26 GH/mm² (overclocked).  This is with no architecture changes just same design and smaller features (the "tock" in Intel's "tick tock" strategy) and scaled out parallel (more cores similar chip dimensions).  There are power issues which prevent Bitfury from achieving the 5 GH/s (420 Mhz) design spec at 55nm.  If they were solved efficiency would be higher.  This is somewhat academic as the 5 GH/s spec was simulated due to the highly efficient use of hand placing.  However hand placing has a lot of pitfalls and Bitfury fell into one making real world performance lower.  "Bitfury28" while having the same general design may not be hand placed.  It is time consuming and more risky.  At 55nm it didn't pay off all the extra work and effort produced a chip which is roughly what we would expect from a cell library design.  Still to be totally exhaustive if the Bitfury had performed as expected it would be 5 GH / (3.8mm * 3.8mm) = 0.34 GH/mm².  With 6x performance improvement that would be ~ 2.00 GH/mm² with a more realistic 4x performance improvement ~ 1.36 GH/mm².

To add to what Aerobatic said about Bitfury being a hand placed chip, this has some implications going forward.  Very few designers hand place large ASICs because of the increased time and cost.  Cell libraries are used to assist with the feature change size, without them you are essentially hand placing a new chip at each process node.  While the design may be exactly the same, the position of each transistor is going to change.  One also needs to consider risk vs reward.  Bitfury's power issues stem from the fact that design was not capable of delivery the power intended, without the intended power the intended clock frequency couldn't be achieved.  The chip "worked" but had to be clocked slower.  It is possible this type of error could have been avoided using a cell library.  So had Bitfury used a cell library maybe they would have got 4 GH per chip, less than what hand placing could do in theory but more than it did in reality.  In mid 2013 Bitfury mistake wasn't fatal.  BFL was not delivering, margins were massive, and other competitors were using much large process nodes.   In 2014 the scenario won't be the same.  A similar mistake would make an offering less attractive than competitors.


*Based on reference design of ~25 GH/s per "H-board" with 16 chips.   A few higher clocked variants (like S-board project) have pushed that to 42-45GH/s.  Assuming I am missing some marginally higher clocked board I optimistically used 3 GH/s as realistic overclock limit.


** Moved this down here because most people probably don't care.  A side note, it may seem I am critical of Bitfury but I am not, I am just interested (obsessed maybe?) with finding good data.  Hell IMHO it is impressive that Bitfury is even still around.  Right at the point where they had perfected their FPGA design (better than anyone else on the same hardware) and were looking to mass produce (summer 2012) BFL began the obviously false campaign of "ASICs in 3 months".  Remember when BFL was going to delivery ASICs by fall 2012?  BFL may be crooked but they are smart.  By offering upgrade value on their FPGA and over-promising ASICs almost a year early it killed the rest of the FPGA market while still keeping BFL FPGA sales alive (they could be upgraded).  Many startups wouldn't have survived seeing their entire market disappear with no revenue potential for a year.  It was unfair, dishonest, and a sucker punch but it wouldn't have surprised me if BFL had killed Bitfury.   Instead the Bitfury team transitioned their FPGA design to a hand placed ASIC in a short period of time, delivered solid performance and did so without the ability to collect (and sit on) preorder money in 2012.  So I am impressed by what Bitfury has done but that doesn't diminish the impressive die efficiency of the Golden Nonce processor.

[-Redacted-]

Just FYI - the main problem with the Bitfury chips, as I understand it, was a "metalization" layout problem - what Punin called "Kappies" (Finnish for pine-cones) appearing where there should have been straight-run bus lines and rounded corners.

Likely a penalty paid for doing a hand layout...

[aerobatic]

So what would a Bitfury28 look like.  With a die shrink we are looking at (55/28)^2 = 4x transistor density.  Lets be generous and say 50% higher clocks are possible so maybe ~6x die efficiency going from 55nm to 28nm. That still puts a purely on paper theoretical Bitfury28 die shrink in the ballpark of say 0.7 GH/mm2 (nominal),  1.26 (overclocked).  Of course this would just be a die shrink with no architecture improvement.  There are obviously power issues which prevent Bitfury55 from acheiving the 5 GH/s (420 Mhz) design spec maybe if they were solved it would have higher die efficiency.

since BitFury's aim was 5 GH, and he achieved approx half that... i think its likely that whatever errors were in the design that prevented it from running at full rate, would get fixed next time around... so, with the die shrink from 55nm to 28nm (4x density of hash cores), and the errors corrected (2x speed), and perhaps some architectural improvements (e.g., fix the daisy chaining which makes it unreliable and hard to scale properly) ...  i would've expected the BitFury28 chip to be at least 8x the current chip's performance...  i.e.: 16 GH (or more), in something thats probably pin compatible with the old one in a low cost qfp as before.   Since its unlikely to run at an 1/8th of the power of the previous one... I'm presuming that the next bitfury28 design will need heatsinks on each chip because the likely power consumption might be in the 8 watt ballpark.

[DeathAndTaxes]

650PH predictions make no sense because of all these reasons and are based on "well, if it keeps going on forever like it has for just the last two months then this is where we'll be" and not on all the things you are bringing up.

FWIW, I end up with very similar numbers when calculating the "end game", using todays BTC value and estimating costs based on HF's chip efficiency and estimated production costs:



The big unknown is how fast we will get there, it will probably take longer than a year, but I have no doubt we will get there if todays BTC value holds.

I like the approach taken but the scale is simply unrealistic. For the sake of the argument lets assume the chips (silicon & packaging) is possible at $0.20 per GH/s ($20 per TH).  Small runs (less than 1,000 wafers) at 28nm are probably more expensive than you so raw silicon is probably close to that and multi-die package with high TDP and high ball count BGA design aren't cheap.  I think >$0.30 per GH/s minimum (excluding NRE and other fixed costs) for chips on a reel is probably more realistic.  However foundry prices are subject to NDA and often opaque so lets make it simple and assume the cost of a chip is $0.00 per GH/s and look at the rest of the system.

For example lets try to guesstimate the balance of the system (everything but the ASIC) on a Sierra.  Here is my guess what is yours?
case: $30 * 1 = $30
watercooling: $80 * 3 = $240
case fans: $10 * 2 = $20
power supplies: $150 * 2 = $300
PCB (both PCB manufacturing and assembly):  $30 x 3 = $90
DC regulators (12V to ~0.8V 200A output ea, 2 per board): $30 x 6 = $180 (probably more KNC uses >$300 per 400 GH/s system)
minor components (connectors, capacitors, etc guesstimate 100+ components per board): $50 x 3 = $150
labor (post PCB assembly, testing, packaging): $50
Balance of system (excluding ASICs) = $1,060 or ~$900 per TH/s.

That is nothing for fixed costs, markup/profit, yield issues, customer support, shipping losses, warranty, etc.

Now let say you could cut that roughly in half to $500 per TH/s.  Your "end game charts" start at half that.  $250 per TH/s might someday be possible but that is a lot of cost to remove from my guesstimate above.  In either case $50 per TH/s is just silly.  Might as well draw a line which shows the end game of $1 per EH/s. $50 per TH/s is $60 per Sierra.  If you think other designs are cheaper to build (excluding chp cost) that would be $25 per KNC Jupiter or $20 per Bitfury-400.  Say I gave you 1.2 TH/s (or 0.5 TH/s, or 0.4 TH/s) of free chips show me how the balance of the system can be made for those prices.  Seriously I would like to see the break down, even a crude guesstimate which gets you into that range.  Pay close attention to DC regulator (VRM) costs and power supplies they are more expensive on a $/GH basis than you might first think.  

Even a generous (I would say unrealistic) scenario of:
8 cent per kWh
20% cooling overhead (PUE 1.2),
no real estate overhead,
$500 per TH hardware
1 year break even for miners
-----------------------------
= network equilibrium at ~225 PH/s.

So it is more like "end game" of <225 PH/s (eventually) at current exchange rate even if my hardware cost is double what is possible.  It kinda shows how silly projections of >650 PH/s in less than a year really are.  Don't feel bad I have seen projections as high as 20 EH/s.

[aerobatic]

I like the approach taken but the scale is simply unrealistic. For the sake of the argument lets assume the chips (silicon & packaging) is possible at $0.20 per GH/s ($20 per TH).  Small runs (less than 1,000 wafers) at 28nm are probably more expensive than you think but lets ignore that and look at the rest of the system.

For example lets try to guesstimate the balance of the system (everything but the ASIC) on a Sierra.  Here is my guess what is yours?
case: $60 * 1 = $60
watercooling: $80 * 3 = $240
case fans: $10 * 2 = $20
power supplies: $150 * 2 = $300
PCB (both PCB manufacturing and assembly):  $30 x 3 = $90
DC regulators (12V to ~0.8V 200A output ea, 2 per board): $30 x 6 = $180 (probably more KNC uses >$300 per 400 GH/s system)
minor components (connectors, capacitors, etc guesstimate 100+ components per board): $50 x 3 = $150
labor (post PCB assembly, testing, packaging): $50
Balance of system (excluding ASICs) = $1,090 or $908 per TH/s.

That is nothing for fixed costs, markup/profit, yield issues, customer support, shipping losses, warranty, etc.

Now say you could cut that in half and had free raw chips that is still >$5,00 per TH/s.  Maybe with enough volume and streamlining you get the cost down to $250 (and assume no yield losses, customer support, profit, etc).  Personnally I doubt it but maybe.  However your charts starts at $250 and go to just silly numbers like $50 per TH/s.  That would be $60 per Sierra equivalent (or $25 per KNC Jupiter).  Say I gave you 1.2 TH/s (or 0.5 TH/s) of free chips show me how the balance of the system would only be $60 or $25.   Pay close attention to DC regulator (VRM) costs and power supplies they are more expensive on a $/GH basis than you might think.

The power axis is misleading because power in = heat out and then it takes more power to remove the heat.  So even an generous (I would say unrealistic) 8 cent per kWh, 20% cooling overhead (PUE 1.2), no real estate overhead, and $500 per TH is more like "end game" of <225 PH/s at current exchange rate.  Kinda shows how silly projections of >650 PH/s in less than a year.

i recently did my own guesstimates of the cost and came to very similar conclusions, though i think you're on the low side for the actual asic costs, but i think you've overestimated the power supply and cooling costs, as they appear to be using relatively off the shelf parts, which will be in volume production and they're not paying retail prices for them.

I think this is very critical, because what it shows is what I've always been saying, that the cost of the system is mostly "other stuff" (psus, cooling, case, pcbs, fans) and not the actual cost of the asics, thus you get enormous cost/performance improvements by having faster asics, as the rest of the system doesn't change in cost at all if the asics went faster!

[Puppet]

I like the approach taken but the scale is simply unrealistic. For the sake of the argument lets assume the chips (silicon & packaging) is possible at $0.20 per GH/s ($20 per TH).  Small runs (less than 1,000 wafers) at 28nm are probably more expensive than you think but lets ignore that and look at the rest of the system.

At the network speeds we are talking about, those wouldnt be small runs. As for the price per chip, these bitcoin asics are no bigger than typical mid/highend mobile SoCs. You've already seen how I came to my estimates, but lets see what ARM has to say on that:

"Our SoC partners might put the first one out on the market with a $15 ASP, and as soon as the second person comes on the market the price drops to $7 and within a month or two, they'll be at $5."

http://www.theinquirer.net/inquirer/news/2303601/arm-reveals-the-mali-t720-gpu-to-help-vendors-get-android-devices-to-market-faster

In all likelihood, mobile socs would have lower yields (far less redundancy), more layers, and far more complicated IO. On the flip side, packaging for low wattage might be easier, but if HF could save money by using seperate chips per die, rather than packaging 4 in one MCM, Im pretty sure they would do it. Really, your low estimate is pretty high. Maybe asic vendors now cant reach that low because they dont have the volume to order directly with the fab, but at some point they will.

For example lets try to guesstimate the balance of the system (everything but the ASIC) on a Sierra.  Here is my guess what is yours?
case: $60 * 1 = $60

Surely you are joking?
Have you seen the cases of KnC? Just 2 flimsy cheats of aluminum that couldnt cost more than $3 in volume. For reference, here is a full fancy looking ATX case for $6
http://www.alibaba.com/product-gs/1016257579/Mini_ATX_Case_Mini_ATX_Tower.html
IT even comes with fans that you price at $10 each

watercooling: $80 * 3 = $240

Again, you have to be kidding. Its not because those prices are common in retail, targeting a select niche market that tolerates such prices, that large volume orders would cost anything like that. $20 tops. Here is a huge radiator, likely the most expensive part,  for $12:
http://www.alibaba.com/product-gs/425032365/Computer_water_cooling_cooling_system.html

case fans: $10 * 2 = $20

You are off by, well, a LOT. Again, for reference first hit on Alibaba, pricing starts at $0.6:
http://www.alibaba.com/product-gs/1059988680/High_tech_120mm_fan_12v_dc.html

power supplies: $150 * 2 = $300
PCB (both PCB manufacturing and assembly):  $30 x 3 = $90
DC regulators (12V to ~0.8V 200A output ea, 2 per board): $30 x 6 = $180 (probably more KNC uses >$300 per 400 GH/s system)
minor components (connectors, capacitors, etc guesstimate 100+ components per board): $50 x 3 = $150
labor (post PCB assembly, testing, packaging): $50

Im not gonna look for prices, but honestly, you are way,WAY off. First crappy 360W PSU I find
http://www.alibaba.com/product-gs/1313266704/360W_Constant_Voltage_12V_Power_Supply.html
Yep, $10 instead of you estimated $150. etc, etc.

I didnt even search for the cheapest, just took one of the first relevant results to give you an idea. Now the above components may be as crap as the infamous BFL power supplies,  might be wrong spec, whatever, but you get the idea. You really confuse retail prices of A brands with what stuff from China costs in volume.

[spiccioli]

I like the approach taken but the scale is simply unrealistic. For the sake of the argument lets assume the chips (silicon & packaging) is possible at $0.20 per GH/s ($20 per TH).  Small runs (less than 1,000 wafers) at 28nm are probably more expensive than you think but lets ignore that and look at the rest of the system.

For example lets try to guesstimate the balance of the system (everything but the ASIC) on a Sierra.  Here is my guess what is yours?
case: $60 * 1 = $60
watercooling: $80 * 3 = $240
case fans: $10 * 2 = $20
power supplies: $150 * 2 = $300
PCB (both PCB manufacturing and assembly):  $30 x 3 = $90
DC regulators (12V to ~0.8V 200A output ea, 2 per board): $30 x 6 = $180 (probably more KNC uses >$300 per 400 GH/s system)
minor components (connectors, capacitors, etc guesstimate 100+ components per board): $50 x 3 = $150
labor (post PCB assembly, testing, packaging): $50
Balance of system (excluding ASICs) = $1,090 or $908 per TH/s.

That is nothing for fixed costs, markup/profit, yield issues, customer support, shipping losses, warranty, etc.

Now say you could cut that in half and had free raw chips that is still >$5,00 per TH/s.  Maybe with enough volume and streamlining you get the cost down to $250 (and assume no yield losses, customer support, profit, etc).  Personnally I doubt it but maybe.  However your charts starts at $250 and go to just silly numbers like $50 per TH/s.  That would be $60 per Sierra equivalent (or $25 per KNC Jupiter).  Say I gave you 1.2 TH/s (or 0.5 TH/s) of free chips show me how the balance of the system would only be $60 or $25.   Pay close attention to DC regulator (VRM) costs and power supplies they are more expensive on a $/GH basis than you might think.

The power axis is misleading because power in = heat out and then it takes more power to remove the heat.  So even an generous (I would say unrealistic) 8 cent per kWh, 20% cooling overhead (PUE 1.2), no real estate overhead, and $500 per TH is more like "end game" of <225 PH/s at current exchange rate.  Kinda shows how silly projections of >650 PH/s in less than a year.

i recently did my own guesstimates of the cost and came to very similar conclusions, though i think you're on the low side for the actual asic costs, but i think you've overestimated the power supply and cooling costs, as they appear to be using relatively off the shelf parts, which will be in volume production and they're not paying retail prices for them.

I think this is very critical, because what it shows is what I've always been saying, that the cost of the system is mostly "other stuff" (psus, cooling, case, pcbs, fans) and not the actual cost of the asics, thus you get enormous cost/performance improvements by having faster asics, as the rest of the system doesn't change in cost at all if the asics went faster!

And to reduce the cost of that "other stuff" bitfury did try to build minimalistic PCBs, but so far has had some problems...

https://bitcointalk.org/index.php?topic=251966.msg3425695#msg3425695

spiccioli

[DeathAndTaxes]

Snipped post to save space.

Well quoting a poorly regulated, low efficiency (70% ouch that is going to hurt power costs) junk PSU, unknown radiator as a substitute for complete sealed waterloop, and "all fans = $0.60" doesn't help your case.  Lets ignore the fact that the site you linked to is notorious for bad prices.  Ask for a firm quote and see how the prices magically change.

Still lets use your imaginary parts:
Junk PSU: 4*10 = $40
Imaginary complete sealed watercooling system: 3 * $20 = $60 (w/ pump, radiator, lines, waterblock, shipped ready to install yeah right.  you pointed out a radiator is $12.  Show me where you can get a pump, reservoir, tubing, connectors, copper waterblock, and assembly for $8 more).
Case: $6 (you don't really believe the listed price on alibaba do you.  Ever asked for quote on a specific model?  Suddenly the $10 special disapears marked up 300% or more)
Fans: 8*$1 ea (its 8 not 2 2 per radiator plus 2 exhaust)

So just these 4 junk components puts you at $74.  A $50 per TH/s target gives you only $60 for an entire system.  Your already overbudget with just the non-electronic components. There is still the ASICs, minor pcb components, pcb manufacturing, pcb assembly, major assembly, and testing.  This also assumes 100% yield, no fixed costs, overhead, taxes, salaries, etc.

Like I said my guestimate of $1000 per system was just a start.  I even said you likely can cut that by 50%.  That is a huge difference from saying you can cut it >95%+ to meet some silly 650 PH/s estimate.

Thanks by your own junk part links you just disproved $50 per TH/s nonsense.  You know it and I know it you just can't admit how utterly silly your projection is.   If you can't source the basic non electronic components (power, case/frame/rack, cooling) for $50 per TH/s it is utterly pointless to show that as a projection.  Moore's law isn't going to make a PSU or case or fan drop 50% in price this year.

[Puppet]

Snipped post to save space.

Well quoting a poorly regulated, low efficiency junk PSU, unknown radiator as a substitute for complete sealed waterloop in China using dubious prices (excluding freight to US) doesn't really help your case.

Still lets use your imaginary parts:
Junk PSU: 4*10 = $40
Imaginary complete sealed watercooling system: 3 * $20 = $60 (w/ pump, radiator, lines, waterblock, shipped ready to install yeah right.  you pointed out a radiator is $12.  Show me where you can get a pump, reservor, tubing, connectors, copper waterblock, and assembly for $8 more).
Case: $6 (you don't really believe the listed price on alibaba do you.  Ever asked for quote on a specific model?  Suddenly the $10 special disapears marked up 300% or more)
Fans: 8*$1 ea (its 8 not 2 2 per radiator plus 2 exhaust)

So that alone using these (and you have to admit) dubious prices and components is $74 before shipping from China.
Your projeciton was prices as low as $50 per TH/s thats $60 per Sierra and you are already overbudget even if we assume your price list is real.  Your already $14 over budget and that is with no ASICs, minor pcb components, pcb manufacturing, pcb assembly, major assembly, and testing.  This also assumes 100% yield.

Like I said my guestimate of $1000 per system was just a start.  I even said you likely can cut that by 50%.  That is a huge difference from saying you can cut it >95%+ to meet some 650 PH/s estimate.

Thanks by your own junk part links you just disproved $50 per TH/s nonsense.  You know it and I know it you just can't admit how utterly silly your projection is.   Hell you can't even get the basic non electronic components (power, case/frame/rack, cooling) for $50 per TH/s much less the entire system.

IM not interested in pricing a Sierra. Frankly thats as silly as doing a BOM analysis on a minirig with an integrated Nexus 7 tablet to make a point that prices per GH couldnt go below $15/GH. Bare bones "designs" that asicminer and bitfury put out, is what future miners will look like. No fancy cases, no water cooling, no integrated tablets, just cheap as chips boards at most in a very basic box of fans. What HF, Cointerra etc are putting out soon is pretty darn nice, but at the prices they can ask for these rigs, there is no reason not to. But in  6 or 12 months when margins evaporate, water cooling  and seasonic PSU's may look as silly as an integrated tablet in a minirig today.

[DeathAndTaxes]

Your "projection" was "based" on Hashfast cost.  It was at the top of your chart.  When asked to back it up you say you aren't interested.  Didn't you ALREADY do such an analysis before coming up with $250 to $50 per TH/s or were those just random numbers you made up?

Still $50 or $100 or $250 per TH/s is just as utterly silly for any vendor, any design, any configuration, anywhere in the world.  It is like saying you will make a clone of an iphone.  Can you do it for $100 maybe, maybe not but if you come in here and say you can mass produce iphone clones for $3.20 ea well it is just laughably dumb.

IIRC the power converters on Bitfury board are about $12 that is in bulk (i.e. 1000+ units).  Just the power regulators are $12 for 40 GH/s. = $30 per TH/s just for one non ASIC component.  Avalon board level BOM is open source (very similar to ASCIMiner who is very hush hush about components) and runs ~$10 per board (excluding ASICs, PCB and assembly).  That's $50 per TH/s for just minor components (crystal, resistors, capacitors, connectors, etc).

You can't get a pcb (no case, fans, cooling, heatsinks, power, host, etc) for the prices you claim even with free ASICs. Just once again to illustrate how silly $50 per TH/s is.  You say Bitfury is the answer to lower cost, at $50 per TH/s that is $2 per board for the overclocked version with higher output power regulators. All PCB production cost, all components, ASICs, assembly, testing, yield losses, etc.  $2 per board.   $2 a board for a 5" by 6" PCB with over 100 components.   I mean it is hard to point that out with a straight face.  Even if your board was $2 power, cooling, open racks, etc have a non-zero cost.

You say quality power isn't important but even at $500 per TH/s & $0.08 per kWh power is about half of the total one year cost.  Using 70% efficient PSU means 30% higher energy cost so robbing Peter to pay Paul.  The junk PSU you listed have horrible voltage regulation and were never designed for high current electronics so what happens when your piece of junk PSU destroys your ASIC.  Is that in your $50 per TH/s target?

[crumbs]

Snipped post to save space.

Well quoting a poorly regulated, low efficiency (70% ouch that is going to hurt power costs) junk PSU, unknown radiator as a substitute for complete sealed waterloop, and "all fans = $0.60" doesn't help your case.  Lets ignore the fact that the site you linked to is notorious for bad prices.  Ask for a firm quote and see how the prices magically change.

Still lets use your imaginary parts:
Junk PSU: 4*10 = $40
Imaginary complete sealed watercooling system: 3 * $20 = $60 (w/ pump, radiator, lines, waterblock, shipped ready to install yeah right.  you pointed out a radiator is $12.  Show me where you can get a pump, reservoir, tubing, connectors, copper waterblock, and assembly for $8 more).
Case: $6 (you don't really believe the listed price on alibaba do you.  Ever asked for quote on a specific model?  Suddenly the $10 special disapears marked up 300% or more)
Fans: 8*$1 ea (its 8 not 2 2 per radiator plus 2 exhaust)

So just these 4 junk components puts you at $74.  A $50 per TH/s target gives you only $60 for an entire system.  Your already overbudget with just the non-electronic components. There is still the ASICs, minor pcb components, pcb manufacturing, pcb assembly, major assembly, and testing.  This also assumes 100% yield, no fixed costs, overhead, taxes, salaries, etc.

Like I said my guestimate of $1000 per system was just a start.  I even said you likely can cut that by 50%.  That is a huge difference from saying you can cut it >95%+ to meet some silly 650 PH/s estimate.

Thanks by your own junk part links you just disproved $50 per TH/s nonsense.  You know it and I know it you just can't admit how utterly silly your projection is.   If you can't source the basic non electronic components (power, case/frame/rack, cooling) for $50 per TH/s it is utterly pointless to show that as a projection.  Moore's law isn't going to make a PSU or case or fan drop 50% in price this year.

Something about this feels like you blew too much money on HF, D&T.
First, let's toss a few unjustified preconceptions:

1.  The chips making up the bulk of hashrate in a year are unlikely to be HF.  I can't predict the tech which will be around in a year's time, but a year ago the hottest thing on the block was FPGA.  How many of those FPGA boards do we need for today's hashrate?

2.  Stop assuming that mining will always be done in the most awkward form factor -- a 4U case with twin PSs & 3 chips, cooled by three radiators & a bevy of fans.
No.
A commercial mining operation can rely on prefab rack units with modular shelving doubling as mounting plates for multiple chip sub-assemblies, complete with quick-disconnect power and cooling.  Water cooling will obviously rely on 1 pump/heat exchanger per rack of ~100 chips.  Heat exchanger could be a water-to-air or water-to-water.  Out in da back.  The cool thing about water is it runs through skinny pipes, nothing special.

3.  Why is everything being shipped to US in your example?  Why not Ukraine, where power costs ~4 cents per KWh?

4.  Why are the PSs multiples of PC power supplies?  You do realize that mining at scale, bigger, less well-regulated PSs could be used, without the pointless multiple voltages/individual cooling/protection/consumer code/etc.?

5.  What makes you think that you know of every chip being developed currently?  Not every company needs to skirt the law & fund their NRE with pre-orders.  Most normal companies don't.  I wouldn't be surprised by unaccountable lumps of hashpower.

If it ever becomes more profitable to mine than to bilk suckers out of their coin, that's exactly what ASIC makers will do.

[Puppet]

Your "projection" was based on Hashfast pricing.  Still $50 or $100 or $250 per TH/s is just as utterly silly for any vendor, any design, anywhere in the world.

I used hashfast to guestimate electricity efficiency and asic costs in general, based on their figures as at that time they were the only 28nm ones with stated die size.  And you've seen the thread where I pulled that chart from, I never even made an attempt to price in the rest of the hardware, I based it solely on the chips and left it up to anyone who wanted to fill out their own assumptions for the non electricity non asic parts for reasons I also explained there. I also never set out to prove 650PH in a year, I didnt come up with that number, and I most certainly didnt come up with that timeline.

But hey, since Im to blame, lets see if we can defend it. Instead of using hashfast, lets substitute it with the Coincraft A1 specs at nominal mode (low power mode might be better, but I cant find a hashrate for it). And lets use some more realistic total system costs and expand the ROI horizon to 2 years. What do we get then?

[RoadStress]

In all likelihood, mobile socs would have lower yields (far less redundancy), more layers, and far more complicated IO. On the flip side, packaging for low wattage might be easier, but if HF could save money by using seperate chips per die, rather than packaging 4 in one MCM, Im pretty sure they would do it. Really, your low estimate is pretty high. Maybe asic vendors now cant reach that low because they dont have the volume to order directly with the fab, but at some point they will.

I haven't been paying much attention to your posts, but after this one i will. So after the many millions of $ cashed by the many ASIC vendors here and after the many scams and lost money + the 2-4(maybe more?) times as much money outside the forum you still think that the volumes will go bigger than this? People running out of money and ROI exponentially lower every day makes you think that mining vendors will live to see BIGGER volumes than ever? WOW...

[Puppet]

you think that mining vendors will live to see BIGGER volumes than ever? WOW...

In terms of hashrate (and thus wafer starts)



Of course unit volumes will skyrocket, almost no one is even shipping 28nm asics yet,  and current prices are at least an order of magnitude above marginal cost.
Dont tell me you thought the network was about to plateau at 5PH ?

[FCTaiChi]

Don't forget there are already manufacturers working on 14-16nm 3d finfets.