50 BTC bet that BFL ASIC will do 350 Mhash/s/Watt or more (ie. Mhash/Joule)

[mrb]

I just opened a bet on our favorite betting platform:

  Butterfly Labs ASIC will achieve 350 Mhash/Joule
  Initial bet: 50 BTC
  http://betsofbitco.in/item?id=665 (full text below)

Many people disagree that BFL will deliver this power efficiency. To these I say: put your money where your mouth is  

At least one ASIC-based Butterfly Labs product will demonstrate 350 Mhash/Joule or more, at stock frequency/voltage, at room temperature, for a sustained period of time of at least 60 minutes, with appropriate current-measuring equipment on the device's power input(s). Power consumption shall be measured by adding up current from all the device's DC inputs (12V jack, USB cable, etc.) It will not be measured "at the wall". The purpose of this bet is to avoid wildly varying efficiencies of power adapters and computer hosts. This bet applies to all ASIC-based products that Butterfly Labs will deliver to the general public. This includes the "SuperComputer" family (Jalapeno, Single SC, Mini Rig SC), as well as possibly any unannounced product that the company will be shipping by June 30, 2013.

This statement is true if, for example, the Single SC achieves 40 Ghash/s at 114 Watt or less (40000/114 = 351 Mhash/Joule) as measured with a clamp meter on its 12V DC input and 5V (USB) input.

This statement is false if all Butterfly Labs products fail to achieve 350 Mhash/Joule, or if Butterfly Labs fails to deliver any product at all.

[1714926359]

1714926359

[1714926359]

1714926359

[1714926359]

1714926359

[1714926359]

1714926359

[mrb]

rsvd

[eldentyrell]

Cool idea, but:

as well as possibly any unannounced product that the company will be shipping by June 30, 2013.

I am not interested in exposing myself to betsofbitcoin's counterparty risk with for nine months.  Or in waiting that long to resolve the bet, to be honest.

Betsofbitcoin desperately needs to offer an "abort" option, i.e. criteria under which the bet will be nullified and all funds returned to bettors.  Then we could have a more reasonable bet that would abort if BFL doesn't deliver by, say, January 1st.

They also really need to switch to intrade-style bid/ask markets.  Since you know how much you stand to lose but do not know how much you stand to win, betsofbitcoin has no predictive power.

[eldentyrell]

Also, you got the bet wrong:

 Butterfly Labs ASIC will achieve 350 Mhash/Joule

BFL is claiming 1400 Mhash/Joule, not 350.  The bet as it is written lets them under-deilver by a factor of 4.  That's a pretty enormous margin of error you're allowing them...

[mrb]

Also, you got the bet wrong: [...]
BFL is claiming 1400 Mhash/Joule, not 350.  The bet as it is written lets them under-deilver by a factor of 4.  That's a pretty enormous margin of error you're allowing them...

No, the bet is not wrong. I have explained multiple times that I believe the Jalapeno will be powered by 2 USB plugs (here, here, and here), that is 700 Mhash/Joule, but I give myself a 2x margin error in this bet, so 350 Mhash/Joule.

However, you have made the claim that ASICs can only be about 4x more efficient than FPGAs at the same process node, so 80 Mhash/Joule. Therefore this bet gives you a greater margin of error: 4.4x.

as well as possibly any unannounced product that the company will be shipping by June 30, 2013.

I am not interested in exposing myself to betsofbitcoin's counterparty risk with for nine months. Or in waiting that long to resolve the bet, to be honest.

You can bet very little if you wish. If you bet only, say, 5 BTC, you would win most of my 50 BTC (assuming you win and no one else bets more on your side).

Betsofbitcoin desperately needs to offer an "abort" option, i.e. criteria under which the bet will be nullified and all funds returned to bettors. Then we could have a more reasonable bet that would abort if BFL doesn't deliver by, say, January 1st.

The bet submitter can define conditions that would nullify the bet. In my case, I decided that if BFL does not deliver, the bet would be lost for me instead of being nullified.

They also really need to switch to intrade-style bid/ask markets.  Since you know how much you stand to lose but do not know how much you stand to win, betsofbitcoin has no predictive power.

You know that if you win, you would at least win the initial amount you bet, plus a variable amount of money. Isn't that sufficient to convince you to bet?

[runeks]

They also really need to switch to intrade-style bid/ask markets.  Since you know how much you stand to lose but do not know how much you stand to win, betsofbitcoin has no predictive power.

You know that if you win, you would at least win the initial amount you bet, plus a variable amount of money. Isn't that sufficient to convince you to bet?

No.

I flip a coin. Will you bet me if I give you 3:1 odds? You should. That would make you money in the long run.

I flip a coin. Will you bet me if I guarantee to pay you back the amount you bet, plus a variable amount of money? You shouldn't. If the "variable amount of money" is less than what you bet, you will lose money in the long run (because your odds would be less than 2:1).

I'd perhaps be willing to place a bet with you, given the odds you provide are enough for me to deem the bet valuable. If I don't know the odds, I have no way of knowing whether the bet is likely to be profitable.

[mrb]

(FYI the exact way the losing bets are distributed amongst the winner is described in How are the losing bets distributed?)

runeks, your analogy is flawed. From eldentyrell's viewpoint, flipping a coin gives him only a 50% probability of winning, however my bet gives him a 100% probability (according to him.)

I find it interesting that he refuses to bet even a symbolic small 5 BTC...

[runeks]

runeks, your analogy is flawed. From eldentyrell's viewpoint, flipping a coin gives him only a 50% probability of winning, however my bet gives him a 100% probability (according to him.)

I find it interesting that he refuses to bet even a symbolic small 5 BTC...

If he really thinks he has no chance of losing then yes, it doesn't make sense not to bet.

I was mostly thinking of myself though. I certainly don't consider the probability of losing to be zero.

[eldentyrell]

I believe the Jalapeno will be powered by 2 USB plugs

So what?  I can run a 30GH/s cluster off of 1200 USB plugs.  That doesn't make it "USB powered".

The USB specification section 6.4.4 and trademark usage requirements are very clear on this issue.  Devices only qualify as "bus powered" if they use less than 500mA.

However, you have made the claim that ASICs can only be about 4x more efficient than FPGAs at the same process node

No, I haven't made that claim.  If I had you would have quoted it.  You seem to be confused.

[mrb]

I don't care what the USB specs says: for all we know, BFL could violate them and have the device pull 1000mA instead of 500mA like some phones do when charging over USB ports of Macs (and some PCs I believe).

The source for your "4x claim" is quoted below. I did not remember that you also said that 10x is at the far end of the possible. So, since 45nm FPGAs do 20 Mhash/J, should I interpret your claim as 200 Mhash/J barely doable with really good engineers on a 45nm ASIC? On top of that, most people (including me) don't believe BFL has the capital to develop at 45nm. 65nm would be more realistic, which would put them at half the efficiency, or 100 Mhash/J.

I still don't understand why you don't take the bet, since I place the bar much higher, at 350 Mhash/J.

In case my intentions aren't clear: I am pushing you to either (1) make you bet money (even a symbolic amount), or (2) force you to concede that 350 Mhash/J is plausible (which would be a reversal of your previous claims, see below)

Another thing that makes the BFL announcement rather
of hard to believe is the very large performance increase
they claim to be able to achieve on an ASIC as compared
to the existing FPGA solutions.

I'd have expected maybe a x3 improvement on - say - the
MH/s/Watts numbers, but the numbers they've announced
are hard to stomach.

I would love for someone really knowledgeable on this topic
(how much more efficient can a chip be made when moving
from FPGA from full custom ASIC).

This.

Just taking your FPGA-tested verilog and pushing it through the Synposys tools will usually get you an ASIC with 4x power improvement.

Working really hard to re-do the design from scratch will get you 8x.  Maybe 10x if you have really good engineers.

A 56x improvement in power consumption is just plain absurd.