Bitfury: "16nm... sales to public start shortly"

[sidehack]

I'd be okay with about $8. I would really like to be able to build a 20-chip board to fit the S1 chassis, with voltage adjustment, and sell it for about $600 per pair. That would pretty much require chips under ten dollars; the lower the better because that $600 can come down.

Regarding chips and who to talk to - I know nothing more than anyone else around here. The only Bitfury contact I've had was with Punin, and I haven't heard anything out of him since about the end of January.

[1715018228]

1715018228

[1715018228]

1715018228

[1715018228]

1715018228

[Mr. Kashif]

I'd be okay with about $8. I would really like to be able to build a 20-chip board to fit the S1 chassis, with voltage adjustment, and sell it for about $600 per pair. That would pretty much require chips under ten dollars; the lower the better because that $600 can come down.

Regarding chips and who to talk to - I know nothing more than anyone else around here. The only Bitfury contact I've had was with Punin, and I haven't heard anything out of him since about the end of January.

I was planning to visit Amsterdam for it.

[PlanetCrypto]

$13 per chip is about twice what it should be.

$13/chip is ~$10/chip over the cost to develop, manufacture, fabricate.
It's typical to ~double the cost to produce a product in setting a MSRP.
By my math, they ought to be around ~$6-$7 each, in any quantity.

These estimates are valid for the first production run.
Follow on production runs have no or significantly reduced embedded RD&D costs, so the price should drop, slightly.
To say $5-$6 /  chip.

Likely there are several reasons supporting the $13/chip MSRP:
1) They are the "only game in town",
2) There are significant benefits from an accounting standpoint for chips used internally priced at $13/chip,
3) Provides justification to "discount" chips supplied to other larger-ish board producers if sales to the community project are charged $13/chip,
4) and lastly, companies are in the business to make money. Maximizing that benefits shareholders. Upper level management reports to and maintains employment from the board of directors. If it came to light that profits were not maximized on a given product line that "official" would be eligible for termination.

As I see it, the community project has 4 choices:
1) "Bite the bullet" and pay the $13/chip,
2) Buy a bunch of S7's, rape the chips off the boards, and let sidehack work his magic on/with efficient board design,
3) A group of community members could band together and create AM version 2, fab a chip (BE400 ) and then produce boards/machines based off that chip,
4) lastly and least likely, a crowd funded community effort to produce a community chip sold at or near cost.

Personally, I like none of the above, for a variety of reasons.

[sidehack]

Number 1 is out, mostly for the reason that in order to not feel like I was screwing people over I'd have to sell miners pretty much at cost, and that doesn't keep me in business.
Number 2 is out because I've tried to raid an S7 and it sucks. Those heatsinks are hard to get off without wrecking the chip, and the chip itself is 0.4mm pin pitch which sucks to work with.
Numbers 3 and 4 are less distasteful but also less likely. Number 3 wouldn't be too bad as long as the people running the place had clear job descriptions and the authority to do their jobs thoroughly - managers do the paperwork and don't screw up the engineers, and engineers have the final say on design. Both sides are essential, but another thing that's essential is not sitting around for a friggin' year with thumbs up butts talking about things that don't matter instead of designing and prototyping a quality product and getting it to market. I like to think about Lockheed back during WW2, when they were asked to design and build a jet fighter (completely from scratch) and deliver a working model within nine months - a task many considered impossible - and they rolled it out ahead of schedule. That's the kind of stuff that can happen when people are allowed to do their jobs without interference.
Another problem is it takes an awful lot of money.

[PlanetCrypto]

Number 1 is out, mostly for the reason that in order to not feel like I was screwing people over I'd have to sell miners pretty much at cost, and that doesn't keep me in business.
Number 2 is out because I've tried to raid an S7 and it sucks. Those heatsinks are hard to get off without wrecking the chip, and the chip itself is 0.4mm pin pitch which sucks to work with.
Numbers 3 and 4 are less distasteful but also less likely. Number 3 wouldn't be too bad as long as the people running the place had clear job descriptions and the authority to do their jobs thoroughly - managers do the paperwork and don't screw up the engineers, and engineers have the final say on design. Both sides are essential, but another thing that's essential is not sitting around for a friggin' year with thumbs up butts talking about things that don't matter instead of designing and prototyping a quality product and getting it to market. I like to think about Lockheed back during WW2, when they were asked to design and build a jet fighter (completely from scratch) and deliver a working model within nine months - a task many considered impossible - and they rolled it out ahead of schedule. That's the kind of stuff that can happen when people are allowed to do their jobs without interference.
Another problem is it takes an awful lot of money.

Fabb'ing a chip to tape out (in qty's of 250,000 minimum) is about 2 million. Don't see that happening anytime soon.

[NotFuzzyWarm]

Crypto hit a couple good points on the $13 pricing with the biggest being BF's internal use accounting: Remember, BitFury as a Corp has several divisions, one being building fixed or 'mobile' Peta-farms to whom the chip div will be selling their product. The chip div makes their markup and the farm building div makes theirs as well. Yeah it's sort of robbing Peter to pay Paul but that's how the money shuffle goes to add up to a very nice bottom-line for the parent company.

Again though, if BitFury really intends to support a Community miner project then screw $13. They know it will not be competing with commercial miner sales so at least give us their best high-volume price or even at-cost. Pretty sure they can do a money shuffle to justify it and still come out ahead.

The other point is that these are 1st production run chips so yeah they will sell at a premium as will the 1st commercial miners using them. As to what price for later runs will be, not sure. the new 16/14nm node pricing cannot be compared to the models for the now very mature higher node sizes. With it just now coming online for boutique chips like miner ASIC's I think TSMC and the very few other foundries capable of producing  16/14nm dies are going to be charging all they can get for quite a while. They have several Tier-1 customers to satisfy using all the production capacity they can throw into that. Everyone else pays a premium to get shoehorned into that.

On the s7 chips.... Sorry about the chip package size Spoiler alert: 100um pitch is now out in the wild as well and our customer in Taiwan is asking us to do our bit to take it under 50um

But as to using bare s7 chips... Once Bitmain releases their 16nm miner, think they would be more open to selling the s7 chips? Just how low of J/GHs can they be taken?

[Gyrsur]

serious question:

where does the price of $13 per chip comes from?

[NotFuzzyWarm]

serious question:

where does the price of $13 per chip comes from?

https://bitcointalk.org/index.php?topic=1291890.msg14633894#msg14633894
Talk to him about it. He seems to be interested in talking to BitFury or has tried to.

[Gyrsur]

serious question:

where does the price of $13 per chip comes from?

https://bitcointalk.org/index.php?topic=1291890.msg14633894#msg14633894
Talk to him about it. He seems to be interested in talking to BitFury or has tried to.

thanks! but we don't know the price per chip if BitFury starts to sell after the halving event. the price should be lower than $13.

[sidehack]

One benefit of BM1385 over A3218 is there's a published datasheet with protocol documentation. However, 0.5mm is about as precise as I can feasibly manufacture without investing probably another $10-15K in equipment. I already have enough trouble with the BM1384 Compacs. I have figured out most of the issues to streamlining, and a chip with pins down two sides instead of all 4 will be easier to wrangle. If I could get BM1385 ASICs would not be ideal, but certainly better than nothing. As it is I'd rather work with Bitfury, or the A3218.

However, probably the best practical efficiency for BM1385 or A3218 is about 0.18J/GH. With the stock setpoint of a Bitfury board being about 60% that power consumption, and the bottom-clock setting about 35% that power consumption, the initial cost of the miner would have to be crazy low to offset the increased operating cost and the long-term viability by comparison is shot to crap.

[2112]

Fabb'ing a chip to tape out (in qty's of 250,000 minimum) is about 2 million. Don't see that happening anytime soon.

This is an example of grammatically correct semantic nonsense.

At "tape-out" the quantity of manufactured chips is zero. At "tape-out" the chip only exist as a "blueprint". It is conceivable that the total cost to tape-out will be zero or nearly zero, if the chip designer can use existing licenses and resources to design the mining chip.

They have several Tier-1 customers to satisfy using all the production capacity they can throw into that. Everyone else pays a premium to get shoehorned into that.

This is an example of salesman bullshit or more precisely somebody's believing and repeating salesman's bullshit.

How do new productions lines come online and get calibrated? By manufacturing super-complex secret designs of Tier-1 customers? No, they are calibrated by manufacturing waferfulls of repeated testing structures for which the fab has a complete detailed precise models to be able to test and calibrate the production line. After testing those test chips are simply scrapped.

Note that in a mining chip is much closer to the fabrication test structure than the typical highly complex chip from a large customer. Therefore mining chips could be profitably manufactured on the production line that is not yet calibrated enough to profitably manufacture very complex designs.

I stress "in theory" because for some reason none of the mining chip designers seems to be able to enter a proper technological partnership with any of the fabricators. This isn't an issue of money, because from my past experience I know of no-budget student projects that had in effect priority access to the new fabrication processes only on the condition of mutual sharing of the design and test data.

If not money, then what is the obstacle? I don't know. Certainly not intellectual property, because SHA256D miner is banal and trivial in the scale of the things that nowadays get manufactured in CMOS.

My guesses go towards some non-technical issue related to the psychology of the mining chips vendors/designers.
 

[sidehack]

Say, question. Do you think it's possible for a 0.1J/GH-order chip to be manufactured on 20/22nm node, and at what cost compared to 14/16nm?

[2112]

Say, question. Do you think it's possible for a 0.1J/GH-order chip to be manufactured on 20/22nm node, and at what cost compared to 14/16nm?

I don't know. The last time I had looked at the actual CMOS transistor models was for BSIM3->BSIM4 transition which only covers processes down to 23/28nm . Edit: I looked at http://www-device.eecs.berkeley.edu/bsim/?page=BSIM4_Arc , this was year 2000. Further work all the way to 2013 extended the coverage to the 20/22nm processes.

My hunch is that the performance of the mining chips could be greatly improved by using a design flow that actually matches the requirements of the mining chip (huge error tolerance margins, ultra deep pipelining) as opposed to the design flows that are the most commonly used on the fringes of the ASIC industry that deal with hit-and-run customers. And that is regardless of the process node and the feature sizes.

It isn't money that prevents the actual knowledgeable people and companies from working with the coin mining vendors. It is something else, but I don't know what.

[sidehack]

Okay. Well I know it's possible to see below 0.2J/GH on 28nm because two or three outfits have done it. From what I understand, 16nm still pulls from something 20nm which actually limits the effective gains from further shrinkage. If it's possible to get about 0.1J/GH out of a more mature (and likely much more cost-effective) process like 20nm, it might be possible to build a compromise miner, with (compared to 16nm) low initial cost and moderate long-term operating cost. Proportionally there's a huge difference between 0.1 and 0.07 but if the 0.1 costs about 1/2 to 1/3 the 0.07 it could still come out ahead, like we're seeing with cheap S7 threatening the viability of higher-cost more-efficient chips.

[2112]

Okay. Well I know it's possible to see below 0.2J/GH on 28nm because two or three outfits have done it. From what I understand, 16nm still pulls from something 20nm which actually limits the effective gains from further shrinkage. If it's possible to get about 0.1J/GH out of a more mature (and likely much more cost-effective) process like 20nm, it might be possible to build a compromise miner, with (compared to 16nm) low initial cost and moderate long-term operating cost. Proportionally there's a huge difference between 0.1 and 0.07 but if the 0.1 costs about 1/2 to 1/3 the 0.07 it could still come out ahead, like we're seeing with cheap S7 threatening the viability of higher-cost more-efficient chips.

You are using the word "cost" incorrectly. The "cost" in the semiconductor manufacturing industry has two parts:

1) https://en.wikipedia.org/wiki/Non-recurring_engineering
2) production costs

I don't know if you are trying to fool yourself or really don't understand the difference.

Only fools or ruthless salesmen or CAD monkeys would claim that total 28nm NRE costs are lower than total 20-14nm NRE costs. This relation may be true for just some step of NRE, like mask-making. The actual proper "NRE design" (conceptual, architectural, simulation and parameter optimization) do not depend on the fabrication node but on the thoroughness of the design.

[Mr. Kashif]

serious question:

where does the price of $13 per chip comes from?

https://bitcointalk.org/index.php?topic=1291890.msg14633894#msg14633894
Talk to him about it. He seems to be interested in talking to BitFury or has tried to.

Here's the proof. I tried talking to them about it and asked if I could give them a visit and the last mail I got from them was that they'll let me know. And till now there's no progress or any response from them.

[sidehack]

What I'm asking about is the expected cost of one chip assuming the design works and reaches mass production. I know the NRE cost is going to be a varying factor not just because NRE is highly variable but because it'd be a fixed initial cost divided amongst an unknown quantity of individual chips. It's not that I don't understand or am trying to fool everyone. I want an idea of the forest, not individual trees.

So, to be more precise than I at first assumed I would have to be to make the question clear to intelligent people seeking to assist rather than get into semantic arguments - let's say the goal is to design a 20nm ASIC that gets 0.1J/GH somewhere in its operating range; could be bottom clock. Give it about 10W expected power dissipation, fairly standard QFN package. Does someone who knows more about semiconductor design think that's possible? If so, assuming we want to produce 1 million ASICs, what would be an expected TOTAL COST combining production costs and NRE, and what would be expected purely for production costs?

[2112]

What I'm asking about is the expected cost of one chip assuming the design works and reaches mass production. I know the NRE cost is going to be a varying factor not just because NRE is highly variable but because it'd be a fixed initial cost divided amongst an unknown quantity of individual chips. It's not that I don't understand or am trying to fool everyone. I want an idea of the forest, not individual trees.

So, to be more precise than I at first assumed I would have to be to make the question clear to intelligent people seeking to assist rather than get into semantic arguments - let's say the goal is to design a 20nm ASIC that gets 0.1J/GH somewhere in its operating range; could be bottom clock. Give it about 10W expected power dissipation, fairly standard QFN package. Does someone who knows more about semiconductor design think that's possible? If so, assuming we want to produce 1 million ASICs, what would be an expected TOTAL COST combining production costs and NRE, and what would be expected purely for production costs?

OK, so you are trying to fool yourself.

This is a gambler's approach to design: state the target and then keep shooting until you hit the target or run out of money/ammunition.

You've posted earlier:

I like to think about Lockheed back during WW2, when they were asked to design and build a jet fighter (completely from scratch) and deliver a working model within nine months - a task many considered impossible - and they rolled it out ahead of schedule. That's the kind of stuff that can happen when people are allowed to do their jobs without interference.
Another problem is it takes an awful lot of money.

Please re-read the book that told the story. This time focus on how much money they had to complete the project, what was their profit/loss model, both for vendor (Lockheed) and customer (US Government). Also analyze when was they key milestone: go/no-go for mass production and what was the percentage of NRE spent before that milestone, that is before they knew if their plane could fly.

Getting back to the mining chip: how much money+time+effort would you/could you spent on simulation of possible designs?

[sidehack]

Right now? Nothing. Not a cent. I could make time, but I currently have no money. Still working on that part. I just want to know if folks who know more than I do think it's possible, and if it is what it might cost to achieve. You're doing a really good job of avoiding what shouldn't be too difficult a question. I know it's going to be an assload of dollars. But comparing to $13 per ASIC, what are we looking at?

[2112]

Right now? Nothing. Not a cent. I could make time, but I currently have no money. Still working on that part. I just want to know if folks who know more than I do think it's possible, and if it is what it might cost to achieve. You're doing a really good job of avoiding what shouldn't be too difficult a question. I know it's going to be an assload of dollars. But comparing to $13 per ASIC, what are we looking at?

Yes, exactly this is your problem: you want to have a guaranteed win at horse-races without investing in horse breeding and having to smell the horse shit.

There's plenty of silver bullet salesmen in the computer industry that will promise you exactly anything that you'll imagine, with no effort, just send them check/coins. Any CAD monkey can drop a SHA256D into a simulator and come up with ballpark figures for a slapdash standard-cell design: Cointerra and Hashfast did just that.

Only in your opinion "shouldn't be too difficult a question". Any serious professional will tell you that it is a non-trivial question.

Basically, you are looking for a slick salesman who will tell you exactly what you like to hear and sprinkle the sales spiel with some technobabble.

Youtube has several comedy videos like that, but they sell turboencabulators to wanna-be motorists.

Edit: Oh, and again a reminder: please don't take every "you" personally. I'm not really writing to "you-sidehack", but to "y'all, youse who support sidehack and other community miner projects".