Topic: tldr: Why isn't KNC able to compete with *55nm bitfury's W/GH? (Read 6660 times)

When you do a full-custom chip design, you actually specify the locations and sizes of the components of the transistors (source, gate, drain, etc)

WTF are you talking about? They never said it was an FPGA copy.  They said it was a standard cell design, and NOT and FPGA copy.


Definitely someone who doesn't know what "RTL" actually stands for, let alone what it means.

I would like to know more about bitfury (and about how one makes it's own transistor, or how smart must one be to go from nothing to best chip in bitcoin in 5 years). What threads can i read? Just his posts here on the forum?

Thanks.

Correct, I since read that elsewhere. Even I cant always be right

bitfury designed their own transistors. Just clarifying.

If I had any doubts, this removes it:
https://www.kncminer.com/userfiles/image/ASIC_PCB.jpg

Thats an altera cyclone FPGA on there. THats what you would use to prototype your design (and pcb).
THe chance that a custom asic would fit, let alone work in the same board as an altera fpga is zero.

Moreover they write underneath that picture:
We will be using these boards to fully validate the entire setup. They will consume the same power, make the same noise level, produce the same heat and run the same RTL code. The only difference will be related to hashing.

Definitely an altera hardcopy implementation.

I believe they advertised it as a FPGA copy, but the tech guy was ripping his eyes out at having to submit it. He wanted to spend more time on it.

Where did they say that and what did they say exactly?
You have to admit, the chip being nearly 4x the size of hashfast, worse power consumption, very fast tape out and promised post tape out implementations, apparently no interest in any tests and done by a company that promotes its hardcopy services. I dont know if they lied or what they said, but if it talks like a duck...

He/she is clearly an uninformed puppet, a simple search would have provided that information.

Are you saying you think they lied when they specifically said it was a standard cell design and not an FPGA conversion?

I think I know now why KnC cant compete with Bitfury on J/GH. And why they are moving so fast. And why they apparently dont feel a need for wafer or chip testing.
They use Orsoc to make the chip. On Orsoc's site I see this:

Retargeting of complex FPGA design from Xilinx Virtex5 into an Altera Hard Copy

http://www.orsoc.se/?page_id=79

KnC's chip is most likely not a custom cell based asic, but a hardcopy V implementation. Didnt KnC work on a FPGA earlier? They would just have used that design as starting point and things can go very fast then. Probably a good move too in this race, but not to win in power efficiency, nor die size efficiency nor cost per unit. But the surest way for time to market and probably more affordable NRE.

chip =/= hash. Oh look at this chip on my desk here [its actually a bottle of blue dragon sweet chilli sauce]. I'm not saying it won't hash, but I want to see it hashing before I consider it a hashing chip.

Oh, here are some pictures of the chips for all the "Where are the chips, they don't exist!!!!" people.

But the actual mining ASIC chip was ready when they started accepting pre-orders. That's the point. I also believe they had the chip hashing before they started taking pre-orders.

It is my thought, based on them saying they were working on a second gen design already, that they probably went standard cell on the first gen in order to get quickly to market with a viable product that they knew they could produce, while reserving the "work of art"  for the next iteration.

I don't know this, of course, it's just my speculation.

But it's the way I would do it. Promise what you are sure you can achieve, and work to blow it out of the water in the meantime.

So far, other than some marketing hiccups out of the gate, KNC has seemed the most professional of the lot, and I'm taking nothing from Bitfury on that. He and his collaborators did deliver the high grade goods ahead of everyone, almost when predicted in fact. That's outstanding. I think he and the KNC crew are worthy competitors in an arena where there are too many dodgy at best operators and a great many straight out scams. Had I been able to raise the funds for my own speculation, I would have bet on KNC to deliver first. Bitfury looked a little scammy to me. I would have bet wrong on delivery, but I still think in the long term KNC has the goods. I also think Bitfury does. This is an interesting game, is it not?

Also, as Bitcoinorama pointed out, KNC have consistently had the mantra of "under promise and over deliver". We only know what they want us to see regarding their actual power consumption. They are playing a close hand, and it may actually be significantly better than we've been told. We will hopefully all know in a couple weeks.

In the meantime, Congratulations to Bitfury. You did a bang up job from left field.

Hypothetical

What chips are you talking about?  I far as I know they don't exist.  I think you talk about specs posted on their sites.
They have no chips.

It makes perfect sense if you know anything about the different ways chips can be designed.

If it didn't actually exist then you can't say if it's power figures are "bad" or "good", they are non-existent.

If it does exist, and it's not as power efficient as bitfury, well, so what?  It's also cheaper, for October delivery, which is all that actually matters as far as profitability is concerned at this point.

The bitfury design is obviously really good - I don't think anyone is denying that. And KnC's chips aren't nearly as efficient as the HashFast/Cointerra chips.

But timing matters far, far, more when it comes to mining profitability then the ultimate 'perfection' of the chip. And KnC has clearly beat those two in time to market.

Punin from bitfurystrikesback.com (EU) and Dave from megabigpower.com (NA) took preorders, one week apart for cscape design hardware using Bitfury chips.

They delivered the majority of the orders within 31 days despite having manufacturing mishaps.

That's not quite comparable. Bitfury started taking preorders after showing a working chip, and then delivered working miners a few months later. KNC took preorders months before even having a chip, still doesn't have a chip, yet promises to deliver working miners in under 2 weeks. The odds of KNC delivering on time are rapidly dropping.

thats a little unfair.  bitfury took pre-orders just like knc did.  bitfury has started shipping now.  and knc is not yet late on its promised shipping dates of end of september so its a little unfair to claim theyre selling squat.  theyre selling pre-orders just like every other asic company including bitfury.  bitfury still has many pre-orders for its october batch, too.   if we get to october and knc hasnt shipped any, then you can crow all you want

I don't get it.  Why you even discussing KNC?

Bitfury sells ASIC chips, KNC sells squat.

Compare them when they actually have a chip to compare.

As it stands now, it is like comparing Bitfury with Dragon ASIC (remember that one?).

I don't get OP's hypocrisy. On one hand he says, on the first post, that he is sick of both camps and their non-existent logic. Then when legitimate arguments form he heavily sides against KNC, repeating the "Equally, each thread is filled with die hard supports who can't see the logic from the fud." that he himself is so apparently against.

This is a problem because the premise of the title of this thread implies that OP is neutral, when as his replies show, he is not. Look for yourself, KNC supporters, bitfury's supporters, or anyone else.

I agree, I got confused and I went to check up their official thread thinking that they really have a working chip!! and in that same official thread people talking about these prototypes as they are in production and just waiting for shipping, I say KNC and all other companies that didn't prove any thing close to a working prototypes are bullish and are in my waiting list ......

Good reply but your action doesn't really reflect what you type. You start a thread about knc chip vs bitfury's and ask why knc is less efficient etc.

People explain to you why that is so and then you just dismiss it off because "knc chips don't exist though". What else are people going to discuss about if you just dismiss it off like that???

You should've just named the thread "tldr: KNC isn't able to compete with *55nm bitfury's w/GH because they don't exist!"

Out of curiosity? It doesn't make sense that 55nm would TROUNCE something that should be 4x more power efficient. I said nevermind because his question is retarded, everyone is talking about KNC vs bitfury, look at the fucking 50000 page thread of KNC people talking about something that also doesn't exist. You can't have it both ways, either it exists and we're discussing it (hence nevermind), or it doesn't exist and its equally shit.

Exactly. Kind of makes you wonder what are the motivations for the OP starting a thread such as this. I'd be interested to hear the answer to plasmoske's question. (or at least a better answer than "nevermind")

well... he sort of does have a point.  You started the discussion comparing their supposed claimed values.  Only to point out one doesn't exist when someone gives the knc chip's advantage.  If we're gonna "stay on topic" all we can do really is hypothetically compare the two which is assumed when you started the thread... just saying.

It seems very unlikely any bitcoin asic provider is doing "full custom" in the sense that they design their own transistors instead of using standard cell libraries, provided either by the fab, or higher level libraries provided by the design house. That to me is the definition of full custom, any other definition is IMO bogus.


Wow, nice numbers, thanks for sharing. That hashfast chip is HUGE. Granted, a bitcoin asic is pretty much a "copy paste" job of a single hashing engine, and I assume they implemented redundancy to cope with single and multiple hash engine defects, but still, such a large chip on such a relatively new process, thats some risk they are taking. I wonder what made them decide on that. For most things, you try to stay under ~100-150mm2 unless you have a very good reason.

Because a one man bitfury team is that much more talented than KNC.

Actually, I have no idea where Ampere and Volt met.  I'm assuming they must have at some point.   

If it doesn't exist then you can't say it's energy efficiency is bad.

What does rolled design mean?  Is it similar to pipelined in CPU architecture?  Not having one part wait for the other but having the parts work in unison like an assembly line?

Full custom has no standard definition.  It is a marketing buzzword these days.  What one companies calls "full custom" another company may not.  A lot depends on the quality of the design and nobody hand places each transistor so custom or not there is some library used, some software is more granular and some is more macro so it is more shades of gray then a black and white distinction.    

That shouldn't take away from Bitfury their design is very efficient. One thing to consider is that Bitfury uses a rolled design and all other designs to date have been unrolled.  Unrolled design has been the conventional wisdom back to the FPGA days, so Bitfury going with a rolled design (which they did on their FPGA as well) may be part of the "magic".  Bitfury likely would have seen commercial success in the FPGA space if BFL hadn't resorted to using underhanded tactics.  BFL announcing ASICs would be available in a few months back in the summer of 2012 killed further demand for FPGAs.  Still Bitfury FPGA design was very efficient compared to competitors so it shouldn't be much surprise that their ASIC implementation would be as well.

Whatever the reason, they use die space very efficiently.  A different way to look at it is how much die space it takes per GH at nominal clockspeed.  Transistor size is directly proportional to process node so I normalized all results to 55nm (native density * (process node/55)^2 ).  KNC & Cointerra not included because they haven't released die size details.  


Note: This should't be taken as an endorsement but more an academical look at how much silicon each design requires to accomplish the same thing if they were fabricated at the same process size.  Everything else being equal less die space per unit of hashing power is better.  It means less silicon cost and compensating for clock frequency and voltage less power usage.  

* Smaller process node is going to have smaller features.  By calculating the change in die size if fabricated at 55nm it provides a better comparison of die efficiency.  Now this isn't a perfect comparison because clock speeds will vary based on process node and how "hard" the manufacturer decides to push the chip.  To do better though requires knowing the clock speed of each chip and the corresponding performance which I don't have or even know if it is available.

I may be wrong but I was under the impression that the BFL chip was not full custom.  There are a few steps between ASIC hard copy and full custom like standard cells.  I am guessing the BFL chip is in the middle of the range.

One might also ask: Why does BFL's 65nm chip consume ~8x as much power per hash as BitFury's 55nm chip?

Both are supposedly Full Custom designs, I believe.

+10 And my You are to stupid too !! 8P

You talk like they've actually delivered something. I'd love KNC to deliver. We need the competition. But until they actually ship something real, their "under promise over deliver" spiel is just marketing BS to lock in pre-order cash to fund their lifestyles just like everyone else.

Yitterbum, I think you're reading me wrong today mate, i'm totally chilled, and i've always enjoyed your comments. I rate Bitfury, he's performed a stella effort considering he is a sole guy, and of course they would need to go for full custom in future. Fact is it's comparing apples to oranges, both went for different solutions with the tech and time available to them, both innovative, both solving a similar problem with an alternative approach, one low cost, full custom, one high cost, but minimal aggrigated risk, both on a finite timeframe, nether products are expected to last forever, but for their window of opportunity, both sufficient solutions, if both succeed. One has, the other we will know very soon.

SIGH nevermind

So if that's what you're saying, why even make a thread about a non-existent chip then???

Absence of evidence isn't evidence of absence.

Bitfury never designed chip in his life before. This one is his first one. Also he says 5-6 years ago he didn't know who are Ampere and Volt and where do they meet each other. This is what he says.

One still doesn't exist though

The perspective is you need to chill, seems like you're taking it personally that someone thinks bitfury's design is better, which is pretty much irrelevant. KnC's are cheaper, come with a case and are pre-configured and plug and play. Obviously a better deal.

But they'll likely need to go for a full custom core to compete with other 28nm chips in the future.

am I missing any news!!! KNC has a chip already ? you are both both talking like you have evidence of a working chip and its results that community is not aware of ( or at least I am not, I haven't been following everything lately )

Right is that not significantly better than their original claims of 250Gh/s and under 1000w? so what perspective do I need exactly? My point still stands they have been pessimistic so as to under promise and over deliver, and not push marketing BS such as Butterfly Labs and the ilk who are claiming their simulated best case as feasible so as to lock in pre-order cash to fund their NRE and future lifestyles.

You really need some perspective, dude. I doubt KnC Jupiter are going to run at less then 500W hashing at 500Gh/s if they're recommending an 850W PSU.

There's nothing wrong with KnC's chips running hotter, since they're actually cheaper.

Are you even aware that KnC have stated only their pessimistic worst case scenarios?! where as Butterfly Labs, Bitfury, Hashfast, Cointerra etc. are the ones that are stating their "predicted" (simulated) best case which is impossible to achieve. That's the difference between marketing and formally trained engineers. In no way do I wish to do Bitfury a disservice as I believe he is a very bright guy and genuinely had the best intentions, and as for Cointerra, they have the credentials, but they are so late to the party they have to make bold claims or no one would fund them. Hashfast are flat out lying about delivery and their Ts&Cs reflect that.

That might be right. BitFury's chip is probably very well optimized.
Keep also in mind that KNC team is more like Friedcat in terms of announcements. They announce what they know is 100% sure.
Other vendors tend to announce fairy tales. Including BitFury to some degree, as you noticed.
For that reason I am expecting that real knc specs might surpass actual figures by some margin.

KnC went with a standard cell design, rather then full custom.  Doing so probably reduced the risks and cost, while reducing performance per watt and per mm². Cointerra and HashFast likely went full custom and have better performance numbers.  

Anyway, it's going to be a long time before power becomes a major cost consideration with ASIC mining so who cares? All that matters at this point is whether or not your hardware ROIs in a month or two.  Cointerra already lowered their prices to $3/Gh for January and we'll likely see tons of price cuts as the market is flooded and difficulty goes through the roof.

agreed. it is underclocked just to insure longevity of the chips. thous not dealing with much customer support for RMA. while allowing them to have more time to do research on future projects. i love the way they do business. also they are the only ones that was able to even ship out at the speed claimed.

Industrial power at $0.02, lol. Why do you think ASICMiner had such a hard time scaling?

My guess is that Bitfury is very skilled in analog chip design and the use of custom transistors, thus enabling him to handcraft a high performance digital chip using only 55nm. On the other hand the KnC/ORSoC people are probably more comfortable with using building blocks (standard cell, FPGA ports, etc.).

No one tested Bitfury chips with external clock (or at least I can't find data) so efficiency may be even better! Rising voltage to get internal clock to get more speed also rises power, and this may be unnecessary. Like in this initial testing. Best reported efficiency is 0.3W/GH!
https://bitcentury.io/blog/initial-testing-of-bitfury-asic

I don't think that logic follows at all.  Bitfury did the math and figured out that for their OWN uses, it was cheaper in the long run to use more chips.  More chips means more complexity and more up-front costs, but if it saves you power in the long run then it is worth it.  Since they are building their own huge mine, they want the best performance per watt as well as the highest reliability.

Saying that KNC will have a lower "ultimate" price point may be true for up front costs... but not lifetime costs. Evaluating lifetime costs is what is going to tell you what your full return on investment will be.

Bitfury purposely underclocked the chip to achieve competitive power specs. That essentially means
they are willing to deliver more total hardware for the equivalent hash rate. To me, that means KnC
will have a lower (ultimate) price point than Bitfury when the pricing war is over.

I'm not sure I understand why you say that Bargraphics has displayed retarded logic. I know him and
I know that he is not a single camp kind of guy. To me, his statement is not coming out in support of
KnC. To me, he is simply stating that in order to have the same power draw for a given hash rate, the
55nm hardware has to be underclocked and over-supplied (comparatively).

Everyone is singing the praises of 28nm architecture, saying that it is the best bang for the buck. All
I see is spending a tremendous amount of upfront money for a huge risk and offloading that risk to a
bunch of pre-ordering ASIC junkies (us). I don't like it. I believe that Bitfury hit the nail squarely on
the head. I believe they will underclock to compete with the 28nm folk and just have an incrementally
higher hardware cost. They may not see the profit that the 28nm folk could ultimately generate, but
they also took A LOT less risk and spent A LOT less OPM (other people's money). Very commendable and
admirable in my opinion.

Lastly, we really don't know what KnC will yield. We only know their estimated power. It could come in
anywhere really. If their chip isn't custom routed like Bitfury's, then is it possible that their performance
differential will only be due to using smaller architecture.

Thats what it looked like, KNC with their 'market leading skills' got it super wrong.

Probably because the 55 nM chips from BitFury are circuitry works of art while the KnC chips are thought to be direct ports of the logic circuits derived from their own less efficient FPGAs.  KnC also chose to work on a more difficult process, so they may experience piss poor initial yields that require huge amounts of voltage to hash.  I would be wary of "predicted" efficiency numbers, as BFL's predictions were off by almost an order of magnitude.  Even BitFury was unable to meet their target, but got pretty close.

   Bitfury asic chip is 55nm but not 65 and BFL is using 65 asic chip.

There we go, the KNCKrew coming out with their retarded logic. You've said NOTHING in that post, literally.

Small thread as its not viable to sift through 100s of pages in each camp. Equally, each thread is filled with die hard supports who can't see the logic from the fud.

The question is, how is bitfury able to achieve 0.8W/GH (confirmed, in hand, 3rd party tested, from WALL) with 655nm, yet KNC is only predicting 1.6W/GH (theoretical, not tested a chip yet, not tested hash rate etc)? Who fucked up? Is KNC lying about their process node?