1GH/s, 20w, $700 (was $500) — Butterflylabs, is it for real? (Part 2) - page 33.

sadpandatech

hero member

Activity: 504

Merit: 500

Quote from: DeathAndTaxes on January 27, 2012, 11:38:15 AM

The question (although it doesn't really matter) is which one.

My curiousity says it matters. Grin

If we take a look at the very impressive chip plot thread. He has damn near fit 3 rings into a 150k @~161MHz = ~241.5MHs. Thats (rings*MHz)*.5 . This leads me to believe BFL could easily use a 102k and fit 2 rings on @ 500MHz+. Which if possible would be ~500MHs per chip, fitting their orginal simulation performance pretty closely.

chip plot thread here; https://bitcointalk.org/index.php?topic=49971.60
I'm not clear if he has them fully unrolled or not. Which would change things a bit obviously.

BFL, just tell us, pleaseeee. It's killing me. Cry

rjk

sr. member

Activity: 448

Merit: 250

1ngldh

Quote from: DeathAndTaxes on January 27, 2012, 11:38:15 AM

Quote from: rjk on January 27, 2012, 11:33:10 AM

Question: Is it completely impossible and/or implausible that they may have developed a custom FPGA?
Another question: Is it plausible that a new FPGA redesigned from scratch with encryption in mind could be better targeted at such processes? My understanding of current FPGAs is that they try as hard as possible to be completely general purpose.

You don't make custom FPGA. You make custom ASICS but that costs couple million dollars and certainly wouldn't get only 10MH/W.

No it almost certainly is (last gen) 60/65nm FPGA that they secure a good deal. The price & wattage fit. The question (although it doesn't really matter) is which one.

I know that, but I am thinking out on a limb. Perhaps they want to become the next Altera Tongue

Which is why I would like to know whether a purpose-built FPGA could be faster for encryption operations than a "general-purpose" FPGA.

From what it sounds like though, it would cost even more than an ASIC to do this.

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: rjk on January 27, 2012, 11:33:10 AM

Question: Is it completely impossible and/or implausible that they may have developed a custom FPGA?
Another question: Is it plausible that a new FPGA redesigned from scratch with encryption in mind could be better targeted at such processes? My understanding of current FPGAs is that they try as hard as possible to be completely general purpose.

You don't make custom FPGA. You make custom ASICS but that costs couple million dollars and certainly wouldn't get only 10MH/W.

No it almost certainly is (last gen) 60/65nm FPGA that they secure a good deal. The price & wattage fit. The question (although it doesn't really matter) is which one.

rjk

sr. member

Activity: 448

Merit: 250

1ngldh

Question: Is it completely impossible and/or implausible that they may have developed a custom FPGA?
Another question: Is it plausible that a new FPGA redesigned from scratch with encryption in mind could be better targeted at such processes? My understanding of current FPGAs is that they try as hard as possible to be completely general purpose.

sadpandatech

hero member

Activity: 504

Merit: 500

I'm under no illusion that they could do it with 40k luts. What happened with the $1200 speculation I popped in there?

Anyhows, what chips do you believe they are using then?

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: sadpandatech on January 27, 2012, 11:14:39 AM

go look thru them again. they start at ~$400... And, safe bet is they are ~$1200 chips that they got from an abandoned project or some such at an extreme discount.

Not with 150K LUTs.

Your $400 part has 40K logic units. To complete 1 hash per clock using the most efficient bitstream known requires 4x as many LUTs. Now it is possible that they have found some magical breakthrough that allows them to perform an SHA-256 hash in 25% of the gates. However if they DID they wouldn't need any Statix, deeply discounted chips, or end of life stock. If hypothetically they could perform 1 hash per clock on 40K LUT budget you could just buy Spartan-6 for $150 put 4 parallel hashers on the bitstream, run it at 200Mhz and get 800MH from a single FPGA board that cost <$200 in bulk.

Assuming magic like that is silly though. So filer your results for units w/ 140K LUT or more & 500 Mhz. Suddenly the prices go way way way way up.

No doubt they got some end of life bulk purchase (or surplus from dead project) but unless they found some chips 99.9% off I doubt they are running at 600Mhz.

sadpandatech

hero member

Activity: 504

Merit: 500

Quote from: DeathAndTaxes on January 27, 2012, 11:12:54 AM

Quote from: sadpandatech on January 27, 2012, 11:09:43 AM

Quote from: DeathAndTaxes on January 27, 2012, 10:58:23 AM

Quote from: DiabloD3 on January 27, 2012, 10:51:06 AM

I'm not saying you're wrong, but, uh, that sounds very wrong.

Care to elaborate?

Remember their original design was 1050MH/s thats 525 MH/s per chip. To achieve that at 1 hash per clock would require a staggering 500Mhz. These aren't CPU or GPU. FPGA tend not to run at that high of a frequency.

Aye, I should have said bitstream instead of software. But the MHz I was refering to is not far fetched. As seen here, from a link I posted a few pages back;
http://components.arrow.com/part/search/1.1v+fpga+stratix+iii

Yeah but those are $4000 parts used for prototyping sASIC and ASIC designs. A pair would be $8000 retail. Even if they got them 90% off you are talking $800 (plus manufacturing, testing, defects, and profit margin).

I guess I should have said. Wider & Slower seems more plausible than ultra fast and narrow.

go look thru them again. they start at ~$400... And, safe bet is they are ~$1200 chips that they got from an abandoned project or some such at an extreme discount.

and, this; EDIT; pic is just to show the heat spreader. this is Strat II and is 1670+ pin.

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: sadpandatech on January 27, 2012, 11:09:43 AM

Quote from: DeathAndTaxes on January 27, 2012, 10:58:23 AM

Quote from: DiabloD3 on January 27, 2012, 10:51:06 AM

I'm not saying you're wrong, but, uh, that sounds very wrong.

Care to elaborate?

Remember their original design was 1050MH/s thats 525 MH/s per chip. To achieve that at 1 hash per clock would require a staggering 500Mhz. These aren't CPU or GPU. FPGA tend not to run at that high of a frequency.

Aye, I should have said bitstream instead of software. But the MHz I was refering to is not far fetched. As seen here, from a link I posted a few pages back;
http://components.arrow.com/part/search/1.1v+fpga+stratix+iii

Yeah but those are $4000 parts used for prototyping sASIC and ASIC designs. A pair would be $8000 retail. Even if they got them 90% off you are talking $800 (plus manufacturing, testing, defects, and profit margin).

I guess I should have said. Wider & Slower seems more plausible than ultra fast and narrow.

sadpandatech

hero member

Activity: 504

Merit: 500

Quote from: DeathAndTaxes on January 27, 2012, 10:58:23 AM

Quote from: DiabloD3 on January 27, 2012, 10:51:06 AM

I'm not saying you're wrong, but, uh, that sounds very wrong.

Care to elaborate?

Remember their original design was 1050MH/s thats 525 MH/s per chip. To achieve that at 1 hash per clock would require a staggering 500Mhz. These aren't CPU or GPU. FPGA tend not to run at that high of a frequency.

Aye, I should have said bitstream instead of software. But the MHz I was refering to is not far fetched. As seen here, from a link I posted a few pages back;
http://components.arrow.com/part/search/1.1v+fpga+stratix+iii

rjk

sr. member

Activity: 448

Merit: 250

1ngldh

Quote from: DeathAndTaxes on January 27, 2012, 10:58:23 AM

FPGA tend not to run at that high of a frequency.

At 30 to 40 watts PER CHIP, I start to wonder though.

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: DiabloD3 on January 27, 2012, 10:51:06 AM

I'm not saying you're wrong, but, uh, that sounds very wrong.

Care to elaborate?

Remember their original design was 1050MH/s thats 525 MH/s per chip. To achieve that at 1 hash per clock would require a staggering >500Mhz. These aren't CPU or GPU. FPGA tend not to run at that high of a frequency.

DiabloD3

legendary

Activity: 1162

Merit: 1000

DiabloMiner author

Quote from: DeathAndTaxes on January 27, 2012, 10:45:00 AM

Quote from: sadpandatech on January 27, 2012, 09:56:09 AM

BFL, are those 500MHz or 600Mhz chips? Will your software run them as fast as can be safely cooled?

FPGA don't have set clock speeds like GPUs. The bitstream controls the clock speed. The same chip with one bitstream may run at 400 Mhz and loaded with a far more intensive bitstream run at 75 Mhz. So you can't make a FPGA run faster via outside software it would require reprogramming the chip w/ a new bitstream.

I don't think there are any FPGA that run at 600 Mhz. More likely they are using a "larger" chip. Spartan 6-150 is used because it takes ~150K LUT to fit a complete unrolled double bitcoin hash logic. Thus 1 hash per clock running at 200 Mhz = 200 MH/s.

If there FPGA have enough LUT to fit 2 complete unrolled hashers then the board would do 4 hashes per clock. 800 MH/s = 200 Mhz.

I'm not saying you're wrong, but, uh, that sounds very wrong.

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: sadpandatech on January 27, 2012, 09:56:09 AM

BFL, are those 500MHz or 600Mhz chips? Will your software run them as fast as can be safely cooled?

FPGA don't have set clock speeds like GPUs. The bitstream controls the clock speed. The same chip with one bitstream may run at 400 Mhz and loaded with a far more intensive bitstream run at 75 Mhz. So you can't make a FPGA run faster via outside software it would require reprogramming the chip w/ a new bitstream.

I don't think there are any FPGA that run at 600 Mhz. More likely they are using a "larger" chip. Spartan 6-150 is used because it takes ~150K LUT to fit a complete unrolled double bitcoin hash logic. Thus 1 hash per clock running at 200 Mhz = 200 MH/s.

If their FPGA have enough LUT to fit 2 complete unrolled hashers then the board would do 4 hashes per clock. 800 MH/s = 200 Mhz.

sadpandatech

hero member

Activity: 504

Merit: 500

Quote from: rjk on January 27, 2012, 09:28:49 AM

Excellent, some REAL porn for once. Fuck boobies, I can fap to this.

aye, very sexy! Those redesigned boards are fuggin hot. I am very anxious to see if testing will show the boards again capable to hashing closer to the projected rates with the new design. We know they should have enough juice for the chips now. If they can keep them cool enough I'd think the rated MHz should be easier to achieve.

BFL, are those 500MHz or 600Mhz chips? Will your software run them as fast as can be safely cooled?

rjk

sr. member

Activity: 448

Merit: 250

1ngldh

Quote from: simonk83 on January 27, 2012, 12:25:06 AM

Excellent, some REAL porn for once. Fuck boobies, I can fap to this.

sadpandatech

hero member

Activity: 504

Merit: 500

Quote from: dirtycat on January 27, 2012, 12:57:50 AM

Quote from: Epoch on January 27, 2012, 12:31:36 AM

Judging by the unmarked heatspreaders, there's no point in removing the heatsinks to find out what chips these are ... there are no markings!

looks like they wire brushed em off

aye, it almost looks like they created the alignment notch themselves too. Maybe to make up for the small divet that is almost completely removed in the process. It's either that or they had custom 65nm chips made some place.(not likely) The world may never know. Well, until they release the units....

DeathAndTaxes

donator

Activity: 1218

Merit: 1079

Gerald Davis

Quote from: antirack on January 27, 2012, 05:55:29 AM

What I somehow don't understand is the motivation of these guys (not only BFL) to make these units and sell them for say $100 to $200 (just a guess out of thin air) profit each. Wouldn't you start mining if you'd have access to hundreds of them? Or do they think the competition is out there so better start selling them?

By that logic shouldn't AMD hardlock their video cards to prevent mining and then just do all the mining themselves.

Some people are gold miners and some people sell the picks and shovels. Historically it has been the vendors who end up profiting more. You also have to consider the time value of money.

Say you have capital to build 100 boards (80 GH). You could either make 10% to 20% (no idea if that is what their margin actually is) in 2 months OR make 200% profit in 4 years at much higher risk. Taking a quick 20% and then using those profits to move forward on another project is usually the wiser choice.

My belief is these chips (based on low cost & high power consumption) are end of life 60/65nm last gen chips. They got a huge discount because they are end of life and the manufacturer wants to clear the product line. If they sell them all off can use the funds to look at a 28nm unit when those parts become more available in mid/late 2012.

bulanula

hero member

Activity: 518

Merit: 500

LOL. So they are probably the real thing but some points still remain.

Why so shady about the chips ? ArtForz will get the chip type in 20 seconds anyway.

Why so many delays ?

What else can it do if the whole thing collapses ?

Costia

newbie

Activity: 28

Merit: 0

Quote from: antirack on January 27, 2012, 05:55:29 AM

Quote from: DiabloD3 on January 27, 2012, 05:54:42 AM

It'll take artforz like 5 minutes to identify it. He already owns like one of every FPGA ever.

Sure, but first they need to ship. I they have collected 1000 orders till the day the ship and they only released pictures with sanded off ICs they have achieved their delay. If they'd put the pictures online with everybody to see what kind of units they are using their competition could go to work right away.

What I somehow don't understand is the motivation of these guys (not only BFL) to make these units and sell them for say $100 to $200 (just a guess out of thin air) profit each. Wouldn't you start mining if you'd have access to hundreds of them? Or do they think the competition is out there so better start selling them?

the profit from mining is speculative. prices of BTC are dropping.
100-200$ profit/unit is a sure thing

antirack

hero member

Activity: 489

Merit: 500

Immersionist

Quote from: DiabloD3 on January 27, 2012, 05:54:42 AM

It'll take artforz like 5 minutes to identify it. He already owns like one of every FPGA ever.

Sure, but first they need to ship. I they have collected 1000 orders till the day the ship and they only released pictures with sanded off ICs they have achieved their delay. If they'd put the pictures online with everybody to see what kind of units they are using their competition could go to work right away.

What I somehow don't understand is the motivation of these guys (not only BFL) to make these units and sell them for say $100 to $200 (just a guess out of thin air) profit each. Wouldn't you start mining if you'd have access to hundreds of them? Or do they think the competition is out there so better start selling them?

Topic: 1GH/s, 20w, $700 (was $500) — Butterflylabs, is it for real? (Part 2) - page 33. (Read 146936 times)