1GH/s, 20w, $500 — Butterflylabs, is it a scam?

[heavyb]

GPUs certainly have resale value, due to their dual purpose. 

But I think that bitcoin price instability is the real killer.  If bitcoin price collapses we are talking two or more year payback period.  And who can predict what will happen after the block reward halves in 2012?

Does it not halve in 2013?

  Depends on how much the block finding rate flucuates between now and then. But, it should be some time between Dec. 2012 and Jan. 2013.

MYAN CALENDAR ENDS IN DECEMBER 2012. More like Earth Reward halves. hehehehe

[1714133148]

1714133148

[1714133148]

1714133148

[1714133148]

1714133148

[1714133148]

1714133148

[Inaba]

Sorry it took so long to get this post up, but I have a few other things that had to be taken care last evening and I didn't get home till late and pretty much went straight to bed.  At any rate, we did a small demo of the hardware last night, here is the test data I used, which I pulled from one of the getwork servers of my pool:

Data: fd90c721557226679bfc01bc971be894ec08137d0f36fd923f822e4743f954da
Merkle: 9d0e5b394ed6ae311a0f61b1

Data: e4f4a3eb23855f185379d5833f0eabb9daee8483e43d39a6a9b3888882bfc0fa
Merkle: 29a9690f4ed6aecc1a0f61b1

Both of these were fresh out of my pool:

The first test should return one nonce of: 22D5485D   
The second should return no nonce value. 

I would ask that someone else validate my test data as well, to be sure what I have is accurate. 

In any case, the first test ran at a hashrate quite a bit lower than projected, but returned the expected nonce on the first round and zero nonces on the second, which was an accurate result.  After this, I did in fact witness a faster test at a speed close to the projected rate, but it was slightly unstable, often returning extra nonces that are invalid.  The explanation I was given for this was plausible and equally plausible that they could work the issues out.  Power consumption for the development unit was substantially higher than projected, but still well within a reasonable amount (as in, it wasn't taking 200 or 300w; it was much, much less). 

Someone asked that I measure the actual chips - they are 30mm^2.

There were some technical difficulties that prevented us from conducting a more thorough test given the time constraints I was under - though BFL was more than willing to take whatever time was needed, even offering to come to my house this evening once they fixed the technical issue that prevented us from doing a comprehensive mining operation - though I am told it's fixed now and ready to go.  But in either case, I feel that the test we conducted showed a POC that adequately demonstrates that at least the hardware does what it's designed to do, if not at the speeds or power consumption stated at the moment. 

Another demo once they get some of the technical issues worked out is planned for the near future (no definite date at the moment, but within a week or so I would think - this is just my speculation) with a fully automated mining client running and submitting work to my live development pool. 

My conclusion is that even if the units were to ship with the lower hashrate I tested and the power consumption I tested, they would still be extremely viable pieces of hardware and are also superior to the currently available public offering(s).  Would they be worth $700 as witnessed?  That would be up to the individual to make that decision, but I personally feel that they are at least within the ballpark of most peoples definition of reasonable.  Any improvements on what I actually saw will increase the value, and from the explanations and technical details I received, I do not see any reason that the final product won't be substantially improved from what I actually tested tonight.

[bittenbob]

Seems to me that they still haven't been able to prove their claims. At a lower hashrate their power consumption was more than stated and the stability issues are very concerning to me. Given that they were unable to run stable at their stated speed and they have had several days to try and figure this out I have my doubts their product will be as advertised. So I have to say FAIL!

[fred0]

Excellent appraisal Inaba.  Thank you for your efforts.

It seems like BFL has accomplished something noteworthy, though not really production quality--yet.

They seem to have the core competency to NOT be a scam.

[DeathAndTaxes]

So a simple case of over promising.  No fancy sASIC, not high end components simply over promising.

$500 was a bait and switch for the $700 price tag.  If power is higher and performance is lower then likely (sorry can't be more specific since our tester was prohibited from giving hard numbers) it is in the same league as our resident FPGA designers.  Maybe a little faster but not a massive performance leap. Also multiple false hashes in one nonce range is very bad. One false hash per nonce range is a 50% rejected share rate.  2 is more like a 66% rejection rate.  3 is a 75% rejection rate.  When 1% reject rate is considered bad it shows just how unstable that board is.

ztex's boards are available at close to $1 per MH if purchased in bulk (licensed production runs not individual retail purchases).
All of the FPGA designs actually available for purchase are getting ~20MH/W.

What made this product "amazing" was getting double those figures (<$0.50 per MH and >50MH/W) if BFL can't delivery that then they never should have promised it.

It would be like ngzhang advertising his board as 400MH.  Technically w/ an optimized enough bitstream it is possible to get 200MH out of a Spartan LX150.  Good thing ngzhang has more integrity than that.  He advertised what he could actually delivery.

Edited for corrections/clarity.

[wormbog]

My conclusion is that even if the units were to ship with the lower hashrate I tested and the power consumption I tested, they would still be extremely viable pieces of hardware and are also superior to the currently available public offering(s).  Would they be worth $700 as witnessed?  That would be up to the individual to make that decision, but I personally feel that they are at least within the ballpark of most peoples definition of reasonable.  Any improvements on what I actually saw will increase the value, and from the explanations and technical details I received, I do not see any reason that the final product won't be substantially improved from what I actually tested tonight.

As a longtime skeptic I can finally say that BF Labs looks to be a real company with a real product. Unfortunately they have a crappy marketing department!

[DeathAndTaxes]

To Inaba:

I know you can't give exact numbers but can you give us more of a ballpark.

When running stable (no false hashes) was performance in the 900 to 1000 MH range or 600 to 800 or <600 MH?

When you say power was more but not 200W well now FPGA on the planet (not even one for 10 years ago) uses 200W.  Hell most CPU don't use more than 200W.

So is "more power" =
still <30W? 
30W - 40W? 
>40W?

Less hashes & more power could mean anything from slightly worse than expected but still decent to worse than products already on sale by other manufacturers.  I know you are prohibited from giving exact numbers but not providing a firmer ballpark range on stats does a disservice to the 3+ FPGA developers who have always supported the Bitcoin community and have real products w/ real verified results.  They have had sales impacted as people wait for results from this magical unicorn.  If you feel you can be firmer but not exact and still meet the spirit of your agreement that would be the right thing to do.

[P4man]

Wait, so you guys are no longer calling them scammers intending only to run off with your money?

It seems you (goat, DnT, etc) were flat wrong on that, and you may well be equally wrong about BFLs inability to fix the performance issues. Particularly since we dont even know how serious they are.

Just sayin'.

[ngzhang]

So a simple case of over promising.  No fancy sASIC, not high end components simply over promising.

$500 was a bait and switch for the $700 price tag.  If power is higher and performance is lower then likely (sorry can't be more specific since our tester was prohibited from giving hard numbers) it is in the same league as our resident FPGA designers.  Maybe a little faster (but multiple false hashes in one nonce range is very bad, i.e. 50%+ server rejection rate) but nothing cutting edge.

ztex's boards are available at close to $1 per MH if purchased in bulk (100 unit licensed run).
ngzhang boards are getting >25MH/W.

What made this product "amazing" was getting nearly double those figures (<$0.50 per MH and >50MH/W) if BFL can't delivery that then they never should have promised it.

It would be like ngzhang advertising his board as 400MH.  Technically w/ an optimized enough bitstream it is possible to get 200MH.  Good thing ngzhang has more integrity than that.  He advertised what he could actually delivery.

oh, sorry sir.
i think i made some math mistake at the "POLL: Miners, do you pay for electricity?  " thread, no 25MH/W, i test them again today, and it's about 18-20MH/w as i say on the main thread : "each board has 2 XC6SLX150 -2FGG484I on it, generates a 360MH/s hashing power. 19.5W on wall power consuming."

360MH pre borad achieved long long ago. actually even 380M is running a long time, but the boards get too hot and earn more power. these days i mainly make efforts on the minning software, i'm totally shit at softwares.

[DeepBit]

ztex's boards are available at close to $1 per MH if purchased in bulk (100 unit licensed run).
ngzhang boards are getting >25MH/W.

It would be like ngzhang advertising his board as 400MH.  Technically w/ an optimized enough bitstream it is possible to get 200MH out of a Spartan LX150.  Good thing ngzhang has more integrity than that.  He advertised what he could actually delivery.

ngzhang is advertising his boards as 360 MH, not 200. May be you were thinking of ztex ?
And ztex is advertising his boards at ~1.6$ per MH, not $1 per MH (100 unit order)

[ngzhang]

ztex's boards are available at close to $1 per MH if purchased in bulk (100 unit licensed run).
ngzhang boards are getting >25MH/W.

It would be like ngzhang advertising his board as 400MH.  Technically w/ an optimized enough bitstream it is possible to get 200MH out of a Spartan LX150.  Good thing ngzhang has more integrity than that.  He advertised what he could actually delivery.

ngzhang is advertising his boards as 360 MH, not 200. May be you were thinking of ztex ?
And ztex is advertising his boards at ~1.6$ per MH, not $1 per MH (100 unit order)

rph is advertising his boards at ~1$ per MH, but not for sell...

[bittenbob]

Wait, so you guys are no longer calling them scammers intending only to run off with your money?

It seems you (goat, DnT, etc) were flat wrong on that, and you may well be equally wrong about BFLs inability to fix the performance issues. Particularly since we dont even know how serious they are.

Just sayin'.

They could still run off with the money seeing as they are not a legal entity so to speak and have not produced any valid business licence. They could be making the boards for themselves and just disappear or maybe they dug themselves a deeper hole than they can  get out of and go bankrupt. If everyone cancels their pre-orders or a significant number do I would have to say I think they would disappear pretty quickly one way or the other.

This is just beginning...

[DeathAndTaxes]

Wait, so you guys are no longer calling them scammers intending only to run off with your money?

It seems you (goat, DnT, etc) were flat wrong on that, and you may well be equally wrong about BFLs inability to fix the performance issues. Particularly since we dont even know how serious they are.

Just sayin'.

Their outlandish performance claims is what led to the belief they were scammers.  A company comming out of nowhere w/ performance claims beyond what is possible w/ FPGA.  That led to speculation on sASICs but the incompatibilty w/ board design as presented, huge upfront cost, the amaturish business mistakes all clashed w/ the kind of investment required to move to that level of integrated circuit production.

They still haven't explained how a 32 card rig box "magically gets" 50x the performance of a single card.  Performance they can't even achieve with 1 card.  Also the card wasn't even mining.  It was simply computing a nonce range.  Real world performance will always be lower than testing a static nonce.


If you feel confident they can fix the "performance" issues I will make a new bet:

BFL will not deliver any product in the next 60 days that meets or exceeds their original claimed specs.
>= 1.05GH/s
<= 19.8W
<= $500

I no longer believe they are running a scam that belief came from performance claims (3x the MH/$ and 2x the MH/W) compared to all other FPGA offerings.  They are simply disreputable business owners vastly overselling what they can deliver.  Their website still indicates 1.05GH/s @ 19.8W guaranteed.  I think in time when they finally delivery a working product to a consumer which is mining in real world conditions we will find out their initial claim 1.05GH @ 19.8W for $500 was bullshit.

[BFL]

Hi Guys,

I'd like to remind everyone that the performance figures Inaba witnessed were on a tuned down debug unit and are not the final performance figures expected for delivery.  Nonetheless, Inaba has confirmed performance is in the class expected.  The reasons for demoing a tuned down unit were clearly stated to Inaba beforehand and to the forum in general.  That is that we are still in development.

The purpose of the pre-release demo was simply to answer the question of 'real' vs 'fake'.   I think most reasonable people would agree that has been resolved.  The next step is final performance and when we complete our dev tuning, you'll have those figures to comment on.  Til then, troll fest!  

Regards,
BFL

[DeathAndTaxes]

ngzhang is advertising his boards as 360 MH, not 200. May be you were thinking of ztex ?
And ztex is advertising his boards at ~1.6$ per MH, not $1 per MH (100 unit order)

ztek offers lower price is the buyer provides all capital.  A licensed production run is $1.26 per MH (in 100 board runs) and $1.02 per MH (in 250 board runs).

ngzhang board uses 2 LX150 chips (theoretical performance is ~200MH/s).  He doesn't advertise the board at 400MH/s because it might someday could acheive that.  He advertises what it can delivery today 360MH/s.   He actually has got it to 380MH/s but doesn't like the heat and power draw.

Shows a supplier w/ integrity.  Working on stability rather than try specsmanship.

[P4man]

BFL will not delivery any product in the next 60 days that meets or exceeds their original claimed specs.
>= 1.05GH/s
<= 19.8W
<= $500

? They never claimed $500 for anything other then a limited pre-order.

I think in time when they finally allow hard numbers we will see they offer nothing special beyond what the legit FPGA designers have acheived.

Maybe you should reread Inaba's post. According to him, the tested unit already represents a significantly better proposition than all the other FPGA. So you think a fixed final shipping product will be worse?

[fred0]

What will the final odds be I wonder?

http://betsofbitco.in/item?id=141

[DeathAndTaxes]

BFL will not delivery any product in the next 60 days that meets or exceeds their original claimed specs.
>= 1.05GH/s
<= 19.8W
<= $500

? They never claimed $500 for anything other then a limited pre-order.

I think in time when they finally allow hard numbers we will see they offer nothing special beyond what the legit FPGA designers have acheived.

Maybe you should reread Inaba's post. According to him, the tested unit already presents a significantly better proposition than all the other FPGA. So you think a fixed final shipping product will be worse?

I would like to see numbers.  Maybe Inaba isn't aware of the performance other boards can achieve when purchased in bulk.  Inaba also seems to indicate the higher performance number (unstable) was the one that made it worthwhile.  And if they never get it stable at that speed?

Also YES I do believe delivered performance will be lower.  Sustained 24/7 operation on a mining pool is always going to be more challenging and lower performance than a theoretical hashrate against a fixed block header.

I would still like to know how a 32 board rig acheived 50x performance?  Maybe I can get one of those performance doublers for my GPU farm.

[DeathAndTaxes]

What will the final odds be I wonder?

http://betsofbitco.in/item?id=141

God bets of bitcoin sucks.  I know my fault for betting but honestly if you are listening b.o.b. please consider moving to an intrade type format.  Had this been a market style bet I would have been buying contracts @ 30 bitcents on the BTC and likely selling them now for 80 bitcents.  Grr.

[P4man]

I gave p4man that bet but with 25% variance.

Clearly they fucked that up.

What makes you think that? I havent read anywhere they were missing their goals by 25%. Where did you read that? For all I know, even as it stands, you will lose the bet with the current state of the software.