[VIDEO] Butterfly Labs (BFL) Bitforce SC ASIC Test

[techwtf]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

[BBQKorv]

Oh my god.

So much fail in one place.

Where is the no significant shipping in May bet?

That would be a really nice bet to make some money. The right horse should be easy to pick 

[Meizirkki]

The only party surprised by higher power usage appears to be BFL itself.. even after they failed to do correct estimates for their FPGAs.

<insert history repetition quote>

[mobodick]

Saw this on unofficial IRC. I'll let you guys pick it apart.

https://www.youtube.com/watch?v=4C4bgho5JSI

Can you embed on here? Is it that flash tag?!

I feel I speak for a lot of the community here when I say...

HOO-BLOODY-RAY!!!

You're hashing prematurely.

[mobodick]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

It looks much more as if they are just testing the controller IC of the board and not the actual ASIC chip.

[willphase]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1. sounds one one of the cores isn't working.
2. single chip on a board - each chip has 16 cores (see image)
3. presumably the test sends a block header (or equivalent) to each core and gets them to iterate over the nonce until the hash < difficulty - they program difficulty in - and the fact that they all came up with the same nonce means they are hashing.  It's the smallest unit of work that an ASIC that does SHA1-256(SHA1-256(data)) would have to do to prove it works.

Basically the video proves that chip works on a board, but only 15 of the 16 cores are hashing, and it uses a lot more power than expected (approx 4x)

Will

[mobodick]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1. sounds one one of the cores isn't working.
2. single chip on a board - each chip has 16 cores (see image)
3. presumably the test sends a block header (or equivalent) to each core and gets them to iterate over the nonce until the hash < difficulty - they program difficulty in - and the fact that they all came up with the same nonce means they are hashing.  It's the smallest unit of work that an ASIC that does SHA1-256(SHA1-256(data)) would have to do to prove it works.

Basically the video proves that chip works on a board, but only 15 of the 16 cores are hashing, and it uses a lot more power than expected (approx 4x)

Will

Aah, thanks for the clarification.

[Dexter770221]

Did I heard "board with extra regulators"? So, problem is in chips not the board. Theydraw to much power. Same story as year ago with SC single. Promised 20W, delivered 80W. Now 60-160, it's progress. With 4th generation maybe they learn how to properly estimate power consumption.

[Frizz23]

Yay! they are not a scam at least. This means they'll eventually ship a product.

fyi: we reached that stage 1.2 month before we shipped. Let's see how they handle their remainder issues.

Based on their track record of failures it will take them at least 5 times of what it took you guys. So we can expect a non-prototype product that works (more or less) within specs in about 1/2 year.

[smoothie]

Josh: We are testing our "chips" and they hash at X rate but should hash at Y rate...could be the boards, the chips, or the clockbuffers...we don't know.

Investor: So when are you guys shipping? Next week? Two weeks?

Josh: We hope/should be shipping in two weeks.

Investor: Will I make profit with my order via mining?

Josh: You "should" profit.

Investor: Yay! I might profit after giving you a ton of money I could have made at least 1000% on in the same time it took you to develop your "asic" device. I love BFL. They are my hero.

 

[Bogart]

Did I heard "board with extra regulators"? So, problem is in chips not the board. Theydraw to much power. Same story as year ago with SC single. Promised 20W, delivered 80W. Now 60-160, it's progress. With 4th generation maybe they learn how to properly estimate power consumption.

BFL Problem: The board or something on it (but we don't think it's the ASIC chips) is drawing 2-3 times as much power as was expected, and we don't know what or why.  But we do know that the voltage regulator can't keep up.

BFL Solution: Add another voltage regulator so that together they'll be more able to handle the increased draw.

[johnyj]

And that 120W draw at run time is only when the chips running at 34Gh/s, so 240W draw for 68Gh/s  

[greyhawk]

"We are still working on the communication protocol"

[techwtf]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1. sounds one one of the cores isn't working.
2. single chip on a board - each chip has 16 cores (see image)
3. presumably the test sends a block header (or equivalent) to each core and gets them to iterate over the nonce until the hash < difficulty - they program difficulty in - and the fact that they all came up with the same nonce means they are hashing.  It's the smallest unit of work that an ASIC that does SHA1-256(SHA1-256(data)) would have to do to prove it works.

Basically the video proves that chip works on a board, but only 15 of the 16 cores are hashing, and it uses a lot more power than expected (approx 4x)

Will

about 2:
One should be shocked if only a single chip is tested.
assume all 16 cores are working, 4GHash (per nonce) * 16 (cores) / 20 (second) = 3.276GH/s per chip, or 205MH/s per hash core.
with this watts shown, you can calculate the efficiency of MHash / Joule by yourself.

the machine should return 8 * 16 = 128 nonces (so the watts can be compared with avalon / asicminer), but why there is only 15...

[johnyj]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1. sounds one one of the cores isn't working.
2. single chip on a board - each chip has 16 cores (see image)
3. presumably the test sends a block header (or equivalent) to each core and gets them to iterate over the nonce until the hash < difficulty - they program difficulty in - and the fact that they all came up with the same nonce means they are hashing.  It's the smallest unit of work that an ASIC that does SHA1-256(SHA1-256(data)) would have to do to prove it works.

Basically the video proves that chip works on a board, but only 15 of the 16 cores are hashing, and it uses a lot more power than expected (approx 4x)

Will

about 2:
One should be shocked if only a single chip is tested.
assume all 16 cores are working, 4GHash (per nonce) * 16 (cores) / 20 (second) = 3.276GH/s per chip, or 205MH/s per hash core.
with this watts shown, you can calculate the efficiency of MHash / Joule by yourself.

the machine should return 8 * 16 = 128 nonces (so the watts can be compared with avalon / asicminer), but why there is only 15...

He mentioned 16 Engines, not 16 cores, some kind of internal function that conbine 8 cores (half of one chip) to make one engine?

All the hashing engines return the same result, so I suppose that the slowest (most unlucky) engine need 18 seconds to finish the hashing function, average could be much less, say 1-2 seconds

[Tehfiend]

Yay! they are not a scam at least. This means they'll eventually ship a product.

fyi: we reached that stage 1.2 month before we shipped. Let's see how they handle their remainder issues.

The primeAsic scammers just released a video that shows just as much proof IMO

[BFL-Engineer]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1) 15 nonce are returned because we run each engine the full range of 0 to 0xFFFFFFFF. They all find the same result, and we read all the engines when they're done.
2) 6 chips on board were tested with the same data. What was shown was the dump of 15 engines on one of the chips.
3) Because the processing range is not distributed among engines, so all the engines go from 0 to 0xFFFFFFFF.



Regards,
Nasser

[Frankie Delaney]

Anyone can tell me:
1 why only 15 nonce like response is returned?
2 what was he tested? single chip? hash core?
3 why it take ~20s to have the power to return idle?

1) 15 nonce are returned because we run each engine the full range of 0 to 0xFFFFFFFF. They all find the same result, and we read all the engines when they're done.
2) 6 chips on board were tested with the same data. What was shown was the dump of 15 engines on one of the chips.
3) Because the processing range is not distributed among engines, so all the engines go from 0 to 0xFFFFFFFF.



Regards,
Nasser

so the test was 15 of 16 engines on 6 of 8 chips? was this done for yeilds or something?

[Frizz23]

What is the die size again?

Package size is 11mm^2.

11mm^2

Die size is 7.5mm^2 if I remember correctly.

20W for 7.5mm^2 is quite a lot of heat to dissipate.

[k9quaint]

Well this is a small step in the right direction.
Maybe after 15 or 20 videos showing steady progress, I might believe they will ship something able to hash at some point in the future.

$30,000 dollars spent on a mini-rig last summer would have bought 5000 bitcoins. Those 5000 BTC would be worth $455,000 today.
BFL has some catching up to do.