Topic: HashFast launches sales of the Baby Jet - page 23. (Read 119648 times)

They'll be fine, but they could've made 3x their money already assuming they handed Barber over the capital when btc was around $100.

Nah, they will be just fine.

Not only miners, but I bet angel investors are pissed too that they didn't just buy btc directly.

I see your point, but Intel being vague on what yield issues were could mean any part of the process is failing. I'm just saying if the company who spends the most money in the world on R&D and chip manufacturing is having issues, i dont see how we can get a <22 bitcoin asic anytime soon.

disclaimer im not an EE  so, just armchair EEing here.

Different types of chips.  If a few gates in a cpu are bad, it's a bad cpu & goes in the dumpster.  If half of a hashing ASIC is bad, it's still perfectly acceptable.  Apples & elephants here.

Don't forget that 22nm and below uses double patterning.  That means 2x as many masks, and fabrication takes 2x as long using the expensive fab equipment 2x as long.  

So people don't take my words out of context.  I never said KNC was giving their employees busy work or that we won't eventually see a sub 28nm processor.  It is just that I strongly (and am willing to eat crow if proven wrong) do not believe the economics make sense for sub 28nm in 2014.  Someday but not in 2014 and probably not in 2015 either (although I am less certain on this).

Some people pointed out that well ASICS have high margins.  Yes they do but the margins are collapsing rapidly.  We have gone from $50 per GH to $10 per GH with Cointerra offer <$3 per GH in Jan, and some lesser names (questionable) offering sub $2 in Feb.  I have no doubt raw chips will see $1 per GH in 2014 and prices will only go lower.  If your cost per mm2 is double that of 28nm the fact that your chip is smaller isn't going to do much good to the bottom line.   It doesn't really makes sense to spend $4M to $8M extra cost to have chips with a HIGHER marginal cost than both your own chips and your competitors.   Eventually 22nm and below WILL be cheaper per mm2 than 28nm but that won't happen for a while.     

Some people have also pointed out that 22nm uses less power and at nominal clock that is correct however it is pretty easy to make silicon use whatever power you need.   Drop the clockrate 20% and you can probably drop the operating voltage 10% to 15%.  The exact amount will depend on the chip design but I am sure all the companies are looking at how low they can undervolt their chips.   Either ship future boards with an adjustable voltage regulator or swap out one resistor and have it operate at a new fixed lower voltage.  Power in is silicon chip is reduced by the square of the voltage reduction so TADA like magic a 10% to 15% drop in voltage means 20% to 30% higher J/GH.  22nm in theory (and reality always falls short or GPU wouldn't need to get bigger and more power hungry) is 30% lower power.   

So 30% lower power with $5M in NRE + 2x (maybe 3x) higher marginal cost or 30% lower power by replacing a $0.10 resistor.  Hmm that is a tough one?  Granted the resistor option means your chips are slower but as pointed out to justify a lower process node the marginal cost is very low.  So take a Sierra as an example (the same thing would apply for a Jupiter or Cointerra rig).   Drop clockrate 20%, voltage 15%, and your use 4 boards instead of 3.  Now you have a marginally faster unit (1.3 TH/s vs 1.2 TH/s) which uses roughly the same power, with same chassis, same power supply, and same boards.

In time the premium on 22nm over 28nm will drop lower and lower and eventually will reach parity with 28nm.   Just like 28nm eventually reached cost parity with 40nm and 40nm reached cost parity with 55nm.  Around that time a respin with an improved design, enhanced functionality, and the benefits of lower power consumption and margin cost makes sense.

Don't believe me how about NVidia.  NVidia estimate for cost parity between 28nm and 22nm is Q1-2015 and cost parity between 22nm and 14n is Q1-2017.   NVidia and AMD won't be using sub 28nm on GPU in 2014.  I mean just think about that for a second if you think 16nm SHA-2 processors are right around the corner.  Two of the biggest companies in processors will be delaying even 22nm until 2015.

Seeing that king of the chips Intel is having yield issues with 14nm, pushing back their broadwell line, i don't see bitcoin asics <22nm anytime soon. I think 2015 is extremely optomistic, and even then pretty pointless  and risky unless we want to mine on our cell phones or something. A lot of miners I know don't have power costs or space concerns, and will gladly buy up any cheap 22nm chips as fast as you spit them out...

It is more about TIME.   The cost depends on complexity but the amount of time it would take would make it a risky bet to tie up your money on considering you have F/X risk during the entire process.   3-4M is a solid bet vs. 2-2.5M for 28

Apple maybe?

We know the cost of 28nm Full Custom is somewhere around 4-6M (Entire Process including Chips) so I imagine slightly below double that possibly.

Does anyone have an educated guesstimate of what the NRE for a 20/16nm asic would be?

However, once margins become that thin if you are the first with 16/20/22nm tech and it is superior even in just power consumption, that company alone would have no competition since no one else would be able to get their NRE back from being second or third.

Hell First might not get it back, but it definitely would be close and secure your place as Industry Leader.

I would imagine that most vendors would look very long and hard about how viable moving to 22nm or lower is in the next couple years. The only real benefit would be the lowered power consumption; 22nm is likely close to as expensive on a per transistor basis as 28nm even outside of NRE. If a 22nm wafer cost the same to process as a 28nm one there might be a cost savings, but that's obviously not the case.
Performance wise, there really are going to be no benefit moving from 28nm to 22nm. Bitcoin is embarrassingly parallel, so there really isn't a performance penalty with using three 100mm^2 28nm dice in place of two 22nm ones at the same area to get the same hashrate.

My gut feeling (not backed up with any market research, BTW) is that the volume of shipments needed to justify the NRE on 20/22nm vs just dropping clocks and voltage on 28nm to reach the same power consumption will likely by much more than the market will be able to bear.

Yeah, KNC gave their engineers (entire workforce) some busywork to keep them from getting underfoot of the assembly monkeys.  It was either that, or having the whole 10Gb company LAN flood with DooM traffic & having to spring for 100Gb
Or they're doing exactly what they said -- working on the next gen, 20/16nm chip.

But for how long? As the network hashrate grows exponentially, the value of these asics drops proportionally. Even the current rise of BTC value is almost meaningless compared to that.  IMO two things can happen; either these companies will start accepting preorders for 20nm chips in the next 3-4 months, or these chips wont materialize.
If they go down the early preorder route, they can lock in the NRE at no risk (other than to the miners who will be screwed again). If they dont, 3-4 months from here, my guess is mining profitability and asic margins will have dropped so low, its probably not worth it anymore.

This is debatable. Profit margins on BTC ASICs are so enormous that price difference between 28nm and  20/16nm probably doesn't matter. Of course we do not have data, but profit margins on KnC chips are probably in thousands of %. Will it stop them to order if 20nm chip is much more expensive?

Well "working on" doesn't mean intend to launch in 2014.  KNC has ASIC engineers. The chips are done unless they intend to say "thanks for all the hard work guys but you are now fired" there is something for them to work on and that is the next version.  We will see in the end but I am still confident no sub 28nm in 2014 (and probably not 2015).  As you know (but often missed by others) TSMC having capacity doesn't mean TSMC having capacity at price point which makes it competitive with existing 28nm designs.

is this your big press release??   How's that swanky substrate?

Hahaha   how is that cointerra chip materializing???

As of two months ago, HashFast was...

Need? Nope. Nor did BFL need preorders to fund Monarch NRE. THe thing is that it locks in preorders at far too high prices, and without those orders its much more questionable if it would make financial sense to spend that money on a new chip.  By the time this chip might come out, prices for 28nm chips would probably be below $100/TH. The new chip might get you higher margins due to its better efficiency, but it wont be anywhere near todays margins, so it will be very difficult to recover the NRE which will also be higher than that of current chips.