Pages:
Author

Topic: CoinTerra Engineering Update: TerraMiner IV Hashing Live (Read 5853 times)

member
Activity: 101
Merit: 10
7uck7m7th


I've traveled the world overclocking competitively...these pictures are from a North American championship I won in 2009...

http://s161.photoBTCucket.com/user/miahallen/slideshow/GOOC2009

I've actually won two NA championships and a world championship  Wink


That is awesome miahallen!! Great stuff!

I would like to see you working on the cutting edge ASIC/Scrypt miners too!

(are you?!)
newbie
Activity: 48
Merit: 0
LOL, thanks, I've thought about it, but I'm not an engineer  Grin


a lot of 'engineers' aren't engineers either
QFT  Roll Eyes   Grin
sr. member
Activity: 434
Merit: 250
Looks nice!  Cheesy
sr. member
Activity: 462
Merit: 250
LOL, thanks, I've thought about it, but I'm not an engineer  Grin


a lot of 'engineers' aren't engineers either
newbie
Activity: 48
Merit: 0
LOL, thanks, I've thought about it, but I'm not an engineer  Grin

In the first picture in that slideshow, the guy wearing the black shirt on the right is Francois Piednoel, one of the lead CPU architects for Intel.  We had some good conversations about CPU design back in the day Smiley
hero member
Activity: 702
Merit: 500
Which is why I mentioned GPUs, which use similar levels (on the high end)  Wink

the same thing.. no gpu runs at the wattage that a hashfast or cointerra asic is going to run at  these things are using hundreds of watts 24/7, and not peaky like a graphics card.  and yet again, no graphics card has as many asics on it, nor as much power requirement as a high end bitcoin miner.  the sheer amount of power that these things will draw compared to a graphics card - significantly more power required.   this isn't just take a graphics card power circuit and multiply by four.  its not that easy.

Actually, the most difficult part of building a reliable VRM is dealing with transitions...so according to what you just stated, ASICs would be easier since they fluctuate less....however, I disagree, I think ASICs power will fluctuate a lot as we know that hashing is a very volatile activity.

Plenty of GPUs run at around 250W TDP...thats about the same as the ASICs in the CoinTerra box...I've owned and overclocked many of them.  I have a little idea of what I'm talking about  Tongue

I've traveled the world overclocking competitively...these pictures are from a North American championship I won in 2009...

http://s161.photobucket.com/user/miahallen/slideshow/GOOC2009

I've actually won two NA championships and a world championship  Wink

with your skill set you should be working for one of these bitcoin mining asic companies !

newbie
Activity: 48
Merit: 0
Which is why I mentioned GPUs, which use similar levels (on the high end)  Wink

the same thing.. no gpu runs at the wattage that a hashfast or cointerra asic is going to run at  these things are using hundreds of watts 24/7, and not peaky like a graphics card.  and yet again, no graphics card has as many asics on it, nor as much power requirement as a high end bitcoin miner.  the sheer amount of power that these things will draw compared to a graphics card - significantly more power required.   this isn't just take a graphics card power circuit and multiply by four.  its not that easy.

Actually, the most difficult part of building a reliable VRM is dealing with transitions...so according to what you just stated, ASICs would be easier since they fluctuate less....however, I disagree, I think ASICs power will fluctuate a lot as we know that hashing is a very volatile activity.

Plenty of GPUs run at around 250W TDP...thats about the same as the ASICs in the CoinTerra box...I've owned and overclocked many of them.  I have a little idea of what I'm talking about  Tongue

I've traveled the world overclocking competitively...these pictures are from a North American championship I won in 2009...

http://s161.photobucket.com/user/miahallen/slideshow/GOOC2009

I've actually won two NA championships and a world championship  Wink
sr. member
Activity: 322
Merit: 250
Supersonic
hero member
Activity: 702
Merit: 500
Which is why I mentioned GPUs, which use similar levels (on the high end)  Wink

the same thing.. no gpu runs at the wattage that a hashfast or cointerra asic is going to run at  these things are using hundreds of watts 24/7, and not peaky like a graphics card.  and yet again, no graphics card has as many asics on it, nor as much power requirement as a high end bitcoin miner.  the sheer amount of power that these things will draw compared to a graphics card - significantly more power required.   this isn't just take a graphics card power circuit and multiply by four.  its not that easy.


newbie
Activity: 48
Merit: 0
Which is why I mentioned GPUs, which use similar levels (on the high end)  Wink
hero member
Activity: 702
Merit: 500
I disagree 100%...designing the power circuits is not easy, yet it's been done thousands of time and there are SO many design ideas to choose from, think about it, every single computer motherboard has a similar DC-DC VRM on it....every single gaming GPU card also has a similar DC-DC VRM.  They're not reinventing the wheel, this is a very well developed science, and they're not having to do anything enormously different than in existing designs.  If they hired the proper engineers, this should be easy-peasy.

The ASIC simulations on the other hand do involve a bit of guess work & luck.  Since they are designing a brand new architecture that has never been done before, they are in unknown territory.  Sure, the chip logic is fairly simple compared to a CPU or GPU, but that doesn't change the fact that there are plenty of unknowns in brand new industry.  You really simply the performance targeting in your description...there are tons of variables that will determine overall performance, the most notable is current leakage...a big issue for IC designers, and it's very difficult to pinpoint the specifics of how this will affect performance.  Basically, finding the right balance between speed, voltage, current, and cooling is a very delicate process.

Think about it...if they were able to find the optimal settings during tape-out as you suggest, then all this tweaking and modifying that they are currently doing would only make things worse.  But in reality, the simulations are never real-world accurate, so tweaking is required to get as close to the theoretical as possible...but that tweaking cannot occur in a lab, it requires working silicon, so that the real world performance can be assessed.

while i agree with you that all pc motherboards have dc/dc converters, they're operating at a very different power level and watts per asic than a bitcoin miner.

an intel cpu is specced for a tdp of 100 watts,.. the extreme ones at 125-130 watts.  if you overclock them maybe they will get to 200 watts.   whereas a bitcoin mining asic will be many more watts... for each chip and in cointerra's case there's four chips on that board, so we're talking a lot of watts.  its a much more extreme version of the power circuit thats in a pc mobo, but agreed, its probably similar components.

i should also mention that an intel cpu requires 1.25-1.5 volts, which is easier and more efficient to deliver from a 12v power supply.. than the 0.7-0.8 volts that a 28nm asic will require.  the bigger volt drop is harder to achieve efficiently.

http://www.overclock.net/t/472058/intel-cpus-maximum-voltages-and-temps
newbie
Activity: 48
Merit: 0
WHAT!?!  No they don't...not for any reasonable customers at least.  Huh

Think about it this way...a company has a chip design and according to the theoretical models, the chip should perform at around 500Gh/s IF conditions A B and C are met.  There is a statistical analysis performed on the chances of delivering on said targets and a probability matrix is compiled.  I'm not a statistics expert, so I'm probably simplifying this a bit.  But once this study is done, the company finds that there is a 20% chance the chip will under-perform by 20% (400GH/s), a 70% chance that the chip will meet the 500GH/s target, and a 10% chance that the chip will over-perform by 20% (600GH/s).

Now, one company decides to announce their product and begin taking pre-orders...what performance target do they publish?  A less business savvy (BFL?) company might not have even performed the statistical analysis, or is looking through green colored glasses, and announces the best case scenario...thinking that somehow the stars will align and the 10% chance of complete success will come true.  He announces the performance specifications of 600GH/s and the rest is history.

Another company come at the situation not counting their chickens before they hatch (KnC?), looks at the reports and thinks about how much more favorable the company will be regarded if they only announces what is relatively certain that they can attain.  After announcing a 400GH/s product, they work hard to achieve & surpass that goal, but they're never under unnecessary pressure to achieve better than a "worst case scenario".

A third company sits somewhere in the middle and thinks that the chances are so good that they'll achieve the most probable result that they announce a 500GH/s part (CoinTerra).  They work hard to achieve the goal, but some of the conditions A, B, or C are unobtainable for reasons outside their control, and the part shows up under-performing their expectations & announcements.

Now, you may think the third option is perfectly fine, but my contention is that the 2nd option is the only honest way to do business...without a lot of luck, because the 1st & 3rd options are are gambles.  Option 1 was obviously the most risky gamble compared to the 3rd...but they were both gambles, and both hands ended coming up short.  When you're playing with your own capitol, gambling may be an ethical option.  But by taking pre-orders and gambling with others capitol the game changes, and the only way to do it ethically (in my opinion) is to place your bets on the most conservative options available (option 2 in my example above).

its not exactly how you describe it.  the process by which each company decides the level of performance is by simulation of the asic after physical design (tapeout) and its pretty much a data science.  if the simulation is done accurately, the final performance should be known for a given instantiation (of the asic).  sure, they could lie.. but i doubt any company sets out to lie (except the truly scammy, or the truly naive).. and i don't think any of the ones we're talking about are either of those.

but the asic isn't the only device in the system, and thats where things get muddy.  in all of these designs, there's a significant analog(ue) component.  the power conversion circuit.  Sam from KnC says that designing this was harder than designing the asic.  (i think he's exaggerating slightly with this bit) but he says that 'designing the asic is the easy part'.  and to some extent i think he has a point.   designing an asic is pretty much a science - its pure solid robust engineering - and can be fully simulated.  designing the high power conversion circuit  that efficiently drops 12 volts from your psu down to less than 1 volt for the asic -  is the part where there's plenty of skill, art, testing and refinement involved.. and its this area where every single asic company has needed more time.  except knc, who, lets be honest.. had two things in their favour.  1.  they over designed their power circuit (originally 8 vrms when they only needed 4) and 2.  they underclaimed the performance so until they could try out the circuit, they didn't know how fast it would actually go (related to how much power they could draw).  in fact, to this day, i think the knc asics will probably go a lot faster than their power conversion circuit will let them.  if they had time (or inclination) to redesign the existing circuit a third time, they would probably have even faster boxes (long before neptune is out).

all of the asic companies are to some extent, power limited - not power at the wall - but power of their conversion circuit.  perhaps it explains why hashfast has shown their asic running off a bench power supply running plenty fast, and yet has spent an inordinate amount of time tweaking their board and still hasn't shipped.  it can't be that the asic is wrong - it seems to work -  ergo, it has to be that the board needs more improvement.   anyhow.  back to cointerra.  since they've shown it running, days after getting one die working, now they have all 16 dies working in a box, with cooling, and power etc.  4 dies each in 4 chips.   thus most of the box works.  now all they need to do is tweak the power circuitry.  they said that in their update.  and to be at 82% of the final goal days after getting a single die working sounds good.  the end game doesn't sound insurmountable to me.  it sounds like a few tweaks here and there and they will be at 100% soon enough.

i think it'd be a shame to criticise cointerra for being this open.  few other companies would give weekly work in progress updates of what their engineers are up to - let alone photos and videos, and I'm quite enjoying seeing some positive steps each week.  Sure, i wish it was happening sooner ... but ive seen real progress each week, and it doesn't sound so far away nor too hard to reach to me, to see a few % improvements to hit their target.   had they just stayed quiet, and waited til they could demo at 100% performance - and then shipped.. we wouldn't be any the wiser. but we would've waited longer between updates.. but they've obviously decided that being open and transparent is their way of keeping customers informed, even if the news they tell customers on a weekly basis won't always be the 'were at 100% and we're shipping now' variety.

I disagree 100%...designing the power circuits is not easy, yet it's been done thousands of time and there are SO many design ideas to choose from, think about it, every single computer motherboard has a similar DC-DC VRM on it....every single gaming GPU card also has a similar DC-DC VRM.  They're not reinventing the wheel, this is a very well developed science, and they're not having to do anything enormously different than in existing designs.  If they hired the proper engineers, this should be easy-peasy.

The ASIC simulations on the other hand do involve a bit of guess work & luck.  Since they are designing a brand new architecture that has never been done before, they are in unknown territory.  Sure, the chip logic is fairly simple compared to a CPU or GPU, but that doesn't change the fact that there are plenty of unknowns in brand new industry.  You really simply the performance targeting in your description...there are tons of variables that will determine overall performance, the most notable is current leakage...a big issue for IC designers, and it's very difficult to pinpoint the specifics of how this will affect performance.  Basically, finding the right balance between speed, voltage, current, and cooling is a very delicate process.

Think about it...if they were able to find the optimal settings during tape-out as you suggest, then all this tweaking and modifying that they are currently doing would only make things worse.  But in reality, the simulations are never real-world accurate, so tweaking is required to get as close to the theoretical as possible...but that tweaking cannot occur in a lab, it requires working silicon, so that the real world performance can be assessed.
newbie
Activity: 48
Merit: 0
I can make it work with 1.62 TH but it would be good to know when the 20% refund is coming... there is one right?
hero member
Activity: 702
Merit: 500
in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

You might want to catch up with the "KNC overclock" thread. I have managed to overclock a 550Gh/s Jupiter to 680Gh/s with no hardware modifications...

its very impressive that you've been able to reverse engineer and defeat their protections...  but I'm sure you agree that its unsupported and probably invalidates your guarantee... but impressive work!

-- Jez


I personally had nothing to do with it and just followed the thread. I'm not sure what protections you are referring too but overclocking involved changing 1 character in text file from 0 to 3 Tongue. Of course I had to use the official voltage tuning they recently added to their firmware to get it stable but there wasn't any "anti-overclocking" protection that I'm aware of. I agree it was impressive work by those who figured out exactly which character to change Cheesy

i was led to believe by Sam who felt it wasn't going to be possible to overclock and they had expended effort to inhibit that.  clearly, ingenuity beats design.
hero member
Activity: 491
Merit: 514
in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

You might want to catch up with the "KNC overclock" thread. I have managed to overclock a 550Gh/s Jupiter to 680Gh/s with no hardware modifications...

its very impressive that you've been able to reverse engineer and defeat their protections...  but I'm sure you agree that its unsupported and probably invalidates your guarantee... but impressive work!

-- Jez


I personally had nothing to do with it and just followed the thread. I'm not sure what protections you are referring too but overclocking involved changing 1 character in text file from 0 to 3 Tongue. Of course I had to use the official voltage tuning they recently added to their firmware to get it stable but there wasn't any "anti-overclocking" protection that I'm aware of. I agree it was impressive work by those who figured out exactly which character to change Cheesy
sr. member
Activity: 462
Merit: 250
in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

You might want to catch up with the "KNC overclock" thread. I have managed to overclock a 550Gh/s Jupiter to 680Gh/s with no hardware modifications...

its very impressive that you've been able to reverse engineer and defeat their protections...  but I'm sure you agree that its unsupported and probably invalidates your guarantee... but impression work!

-- Jez


where ya been? They shipped the first 4VRM miners with over clocking firmware by mistake...  so we all knew they could run much faster once the later firmwares ironed out the hardware errors



hero member
Activity: 702
Merit: 500
in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

You might want to catch up with the "KNC overclock" thread. I have managed to overclock a 550Gh/s Jupiter to 680Gh/s with no hardware modifications...

its very impressive that you've been able to reverse engineer and defeat their protections...  but I'm sure you agree that its unsupported and probably invalidates your guarantee... but impressive work!

-- Jez
hero member
Activity: 702
Merit: 500
WHAT!?!  No they don't...not for any reasonable customers at least.  Huh

Think about it this way...a company has a chip design and according to the theoretical models, the chip should perform at around 500Gh/s IF conditions A B and C are met.  There is a statistical analysis performed on the chances of delivering on said targets and a probability matrix is compiled.  I'm not a statistics expert, so I'm probably simplifying this a bit.  But once this study is done, the company finds that there is a 20% chance the chip will under-perform by 20% (400GH/s), a 70% chance that the chip will meet the 500GH/s target, and a 10% chance that the chip will over-perform by 20% (600GH/s).

Now, one company decides to announce their product and begin taking pre-orders...what performance target do they publish?  A less business savvy (BFL?) company might not have even performed the statistical analysis, or is looking through green colored glasses, and announces the best case scenario...thinking that somehow the stars will align and the 10% chance of complete success will come true.  He announces the performance specifications of 600GH/s and the rest is history.

Another company come at the situation not counting their chickens before they hatch (KnC?), looks at the reports and thinks about how much more favorable the company will be regarded if they only announces what is relatively certain that they can attain.  After announcing a 400GH/s product, they work hard to achieve & surpass that goal, but they're never under unnecessary pressure to achieve better than a "worst case scenario".

A third company sits somewhere in the middle and thinks that the chances are so good that they'll achieve the most probable result that they announce a 500GH/s part (CoinTerra).  They work hard to achieve the goal, but some of the conditions A, B, or C are unobtainable for reasons outside their control, and the part shows up under-performing their expectations & announcements.

Now, you may think the third option is perfectly fine, but my contention is that the 2nd option is the only honest way to do business...without a lot of luck, because the 1st & 3rd options are are gambles.  Option 1 was obviously the most risky gamble compared to the 3rd...but they were both gambles, and both hands ended coming up short.  When you're playing with your own capitol, gambling may be an ethical option.  But by taking pre-orders and gambling with others capitol the game changes, and the only way to do it ethically (in my opinion) is to place your bets on the most conservative options available (option 2 in my example above).

its not exactly how you describe it.  the process by which each company decides the level of performance is by simulation of the asic after physical design (tapeout) and its pretty much a data science.  if the simulation is done accurately, the final performance should be known for a given instantiation (of the asic).  sure, they could lie.. but i doubt any company sets out to lie (except the truly scammy, or the truly naive).. and i don't think any of the ones we're talking about are either of those.

but the asic isn't the only device in the system, and thats where things get muddy.  in all of these designs, there's a significant analog(ue) component.  the power conversion circuit.  Sam from KnC says that designing this was harder than designing the asic.  (i think he's exaggerating slightly with this bit) but he says that 'designing the asic is the easy part'.  and to some extent i think he has a point.   designing an asic is pretty much a science - its pure solid robust engineering - and can be fully simulated.  designing the high power conversion circuit  that efficiently drops 12 volts from your psu down to less than 1 volt for the asic -  is the part where there's plenty of skill, art, testing and refinement involved.. and its this area where every single asic company has needed more time.  except knc, who, lets be honest.. had two things in their favour.  1.  they over designed their power circuit (originally 8 vrms when they only needed 4) and 2.  they underclaimed the performance so until they could try out the circuit, they didn't know how fast it would actually go (related to how much power they could draw).  in fact, to this day, i think the knc asics will probably go a lot faster than their power conversion circuit will let them.  if they had time (or inclination) to redesign the existing circuit a third time, they would probably have even faster boxes (long before neptune is out).

all of the asic companies are to some extent, power limited - not power at the wall - but power of their conversion circuit.  perhaps it explains why hashfast has shown their asic running off a bench power supply running plenty fast, and yet has spent an inordinate amount of time tweaking their board and still hasn't shipped.  it can't be that the asic is wrong - it seems to work -  ergo, it has to be that the board needs more improvement.   anyhow.  back to cointerra.  since they've shown it running, days after getting one die working, now they have all 16 dies working in a box, with cooling, and power etc.  4 dies each in 4 chips.   thus most of the box works.  now all they need to do is tweak the power circuitry.  they said that in their update.  and to be at 82% of the final goal days after getting a single die working sounds good.  the end game doesn't sound insurmountable to me.  it sounds like a few tweaks here and there and they will be at 100% soon enough.

i think it'd be a shame to criticise cointerra for being this open.  few other companies would give weekly work in progress updates of what their engineers are up to - let alone photos and videos, and I'm quite enjoying seeing some positive steps each week.  Sure, i wish it was happening sooner ... but ive seen real progress each week, and it doesn't sound so far away nor too hard to reach to me, to see a few % improvements to hit their target.   had they just stayed quiet, and waited til they could demo at 100% performance - and then shipped.. we wouldn't be any the wiser. but we would've waited longer between updates.. but they've obviously decided that being open and transparent is their way of keeping customers informed, even if the news they tell customers on a weekly basis won't always be the 'were at 100% and we're shipping now' variety.


newbie
Activity: 48
Merit: 0
Yes, I'm a little disappointed.  I was hoping (especially since they missed their target shipping date) that the 2TH/s figure was a very conservative estimate.  But it appears it's going to be a challenge to get it there which means if they are able to achieve it, the performance/watt target my get pushed (as we've seen from other vendors), and consequently it means that there's much less potential for overclocking.  I was hoping this news would come with a surprise announcement that they were already able to exceed their goal by 20% (2.4TH/s) or more much like KnC did.  As the two top manufacturer's in the market, KnC is their primary competition in my opinion, and based on the results from both companies, I think I'm a bigger fan of KnC and the philosophy by which they operate.

I'm invested in both as I own a 1st batch hosted KnC Jupiter, and I have two (2) Terrahash IV (early JAN) miners coming my way soon as well.  So, while I'm also very happy to see a working product, and they've done a great job overall...it seems to me they may have been a bit hasty with their performance predictions, and that's a little discouraging.  This isn't a bashing rant or anything...I'm very patient, and I'm confident this will be the best product on the market for several months (which is why I ordered two).

hope you forgive the length of this answer...

its premature to know whether it will overclock (or even if cointerra will allow it).   in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

its possible knc deliberately underquoted their specs so that they've could pleasantly surprise their customers with 'more' when they deliver - sounds great if you're a customer (and i am).. and is fantastic to hear.. but when you make it your motto, and the customers expect it (and require it).. then you're setting them up for a potential disappointment if you don't hit the same amount of over delivered performance each and every time you deliver a new product.    if you make it your motto and stake your reputation to 'under promise and over deliver'... its almost a double edged sword.  on the one hand, you gain a fantastic reputation because people are expecting you to over deliver.. but on the other hand, you had better over deliver by enough..!  by as big a margin as you did last time, or they might not be happy about the bonus!  it'll get harder each time as you approach the limits of power, cooling, technology, etc.

its difficult to engineer something that will always have power to spare.  many engineers would even say that they already know in advance how fast it will go and aren't waiting 'to try it' before they find out how fast it will go.   knc had a fantastic opportunity by entering a market with a properly engineered, high tech solution in a market full of amateurs and lousy engineering.  but the market has changed and its no longer full of bad asic designs and amateurs.   the first time you do it, you can under-quote  because you know your chip is wayy better than everyone else.  but when the market becomes more crowded, its a lot harder to compete.   in the competitive world of pre-ordered miners, if you underquote what your box will do, you will look uncompetitive compared to the people who aren't underquoting or who are just naively optimistic (are you listening Ken?) and lose out to the people who are prepared to bend the truth (some others not to be named).  I'm not saying this has ever happened...!

.. but i somewhat agree with you in that with most of the 'high power' chip people - kncminer, hashfast, cointerra and even bitmine etc... that the power delivery seems to have become more of an issue than any of them could've possibly imagined and the way they've squeezed out more performance is by improving their power circuits.  ironic that the asic chips in most cases have worked first time (what are the chances!?) -  whereas the power input - good old fashioned analog circuit design - has become such a 'dark art' to be solved, and required lots of tweaks to find the best solution.   in some cases, like hashfast, this has taken significant amounts of time.

in cointerra's case (much like knc did) they seem to have over engineered some aspects of the design - the available power & cooling capacity... which should allow them a little room to tweak.  whether there's extra margin for over clocking, i doubt it.   knc doesn't let anyone overclock, for a good reason.  these things are built on the limit of the rest of the circuitry.    but I'm ok with cointerra taking a bit of extra time to deliver the performance they promised (especially as i am expecting a january batch, and they're not even a day late for my orders, yet, and i hope they arent).

if this was a performance car... its sold mostly on its performance compared to its peers.  you can't really get away with under declaring its performance (to delight your customers with a surprise extra bonus) and still sell your high performance cars as your competitors will be saying theirs is faster, even if it won't be.   knc did it once.  they over delivered by a huge margin.  that could easily be their motto... but now they have to over-deliver each and every time, which is really difficult to keep doing, especially when others aren't doing that.   as we approach the limits of power & cooling in a home system, i don't think significant over delivery is a viable future option.  in an ideal world... you engineer for what it can do, and do your best to deliver it in a reasonable time.

Quote
knc did it once.  they over delivered by a huge margin.  that could easily be their motto... but now they have to over-deliver each and every time

WHAT!?!  No they don't...not for any reasonable customers at least.  Huh

Think about it this way...a company has a chip design and according to the theoretical models, the chip should perform at around 500Gh/s IF conditions A B and C are met.  There is a statistical analysis performed on the chances of delivering on said targets and a probability matrix is compiled.  I'm not a statistics expert, so I'm probably simplifying this a bit.  But once this study is done, the company finds that there is a 20% chance the chip will under-perform by 20% (400GH/s), a 70% chance that the chip will meet the 500GH/s target, and a 10% chance that the chip will over-perform by 20% (600GH/s).

Now, one company decides to announce their product and begin taking pre-orders...what performance target do they publish?  A less business savvy (BFL?) company might not have even performed the statistical analysis, or is looking through green colored glasses, and announces the best case scenario...thinking that somehow the stars will align and the 10% chance of complete success will come true.  He announces the performance specifications of 600GH/s and the rest is history.

Another company come at the situation not counting their chickens before they hatch (KnC?), looks at the reports and thinks about how much more favorable the company will be regarded if they only announces what is relatively certain that they can attain.  After announcing a 400GH/s product, they work hard to achieve & surpass that goal, but they're never under unnecessary pressure to achieve better than a "worst case scenario".

A third company sits somewhere in the middle and thinks that the chances are so good that they'll achieve the most probable result that they announce a 500GH/s part (CoinTerra).  They work hard to achieve the goal, but some of the conditions A, B, or C are unobtainable for reasons outside their control, and the part shows up under-performing their expectations & announcements.

Now, you may think the third option is perfectly fine, but my contention is that the 2nd option is the only honest way to do business...without a lot of luck, because the 1st & 3rd options are are gambles.  Option 1 was obviously the most risky gamble compared to the 3rd...but they were both gambles, and both hands ended coming up short.  When you're playing with your own capitol, gambling may be an ethical option.  But by taking pre-orders and gambling with others capitol the game changes, and the only way to do it ethically (in my opinion) is to place your bets on the most conservative options available (option 2 in my example above).
hero member
Activity: 491
Merit: 514
in some ways knc is responsible for setting unrealistic expectations in their customers but one thing they never did is allow their customers to overclock.  not only is it forbidden but there's extra circuitry in every knc box to make it extremely hard to do.

You might want to catch up with the "KNC overclock" thread. I have managed to overclock a 550Gh/s Jupiter to 680Gh/s with no hardware modifications...
Pages:
Jump to: