Pages:
Author

Topic: BlockBurner LLC - Crucible FPGA Scrypt Miner - Announcement Aug-19 - page 6. (Read 42404 times)

newbie
Activity: 22
Merit: 0
1. Do you think the market and community is ready for FPGA Litecoin?
Market, maybe. Community, YES! If you can pull better numbers per joule and come in cheaper on hardware cost vs gpu, people will eat em up!

This part (cheaper than GPU's for a given hash rate) is unlikely.  When it comes to scrypt(1024,1,1), a modern Radeon GPU (prior to the 7xxx series, anyway) very nearly is an ASIC already optimized for scrypt, and designed at optimal process node and with economies of scale of being a consumer product.  I've already given it a go with the Xilinx Artix-7 FPGA's surrounded by large quantities of DDR3 and tried a variety of approaches across the TMTO spectrum for the known methods of calculating scrypt, ranging from no use of external memory at all and pipelining the entire calculation, to replicating the way cgminer does it in OpenCL (with various lookup gaps).  In the end, the best use of the prototype hardware was actually to mine the heck out of Yacoin, since scrypt+chacha20/8+keccak(N,1,1) with N=32 (as it currently is for Yacoin until tomorrow) was almost trivial to optimize for FPGA's:

https://bitcointalksearch.org/topic/m.2127307

For the probable FPGA vs. GPU cost relationship to change, someone needs to find and disclose a method of further shortcutting the already-known TMTO of scrypt(1024,1,1).  And at that point, whatever that optimization happens to be, is very likely going to be equally applicable to OpenCL to speed up GPU processing of scrypt too.

What I have yet to see is either the BlockBurner team or jasinlee say "Oh yeah, we're genius cryptanalysts and found ways that scrypt(1024,1,1) can be calculated much faster and/or with significantly less logic than anyone else has figured out how to."

What is the frequency and total data bits width of your DDR3  used?
full member
Activity: 210
Merit: 100
That is a very well thought out and technical rebuttal to my pipe dream. Thanks for ruining christmas.

My (unskilled) thinking was that scrypt is ram heavy, ram is pretty cheap these days. Math, on the other hand, is not my strong suit. Thanks for the perspective and publically abusing limitless in that thread you linked to. Wink

dram size is cheap, sram is expensive, and bandwidth is very expensive.
GPU can easily have more than 300GB/s bandwidth, but FPGA less than 30GB/s.
IF there exists a FPGA can reach a bandwidth >100GB/s, I think we can do it.

There are FPGAs that can do it.  The problem is they cost 5k and up. 
newbie
Activity: 22
Merit: 0
That is a very well thought out and technical rebuttal to my pipe dream. Thanks for ruining christmas.

My (unskilled) thinking was that scrypt is ram heavy, ram is pretty cheap these days. Math, on the other hand, is not my strong suit. Thanks for the perspective and publically abusing limitless in that thread you linked to. Wink

dram size is cheap, sram is expensive, and bandwidth is very expensive.
GPU can easily have more than 300GB/s bandwidth, but FPGA less than 30GB/s.
IF there exists a FPGA can reach a bandwidth >100GB/s, I think we can do it.
full member
Activity: 210
Merit: 100
1. Do you think the market and community is ready for FPGA Litecoin?
Market, maybe. Community, YES! If you can pull better numbers per joule and come in cheaper on hardware cost vs gpu, people will eat em up!

This part (cheaper than GPU's for a given hash rate) is unlikely.  When it comes to scrypt(1024,1,1), a modern Radeon GPU (prior to the 7xxx series, anyway) very nearly is an ASIC already optimized for scrypt, and designed at optimal process node and with economies of scale of being a consumer product.  I've already given it a go with the Xilinx Artix-7 FPGA's surrounded by large quantities of DDR3 and tried a variety of approaches across the TMTO spectrum for the known methods of calculating scrypt, ranging from no use of external memory at all and pipelining the entire calculation, to replicating the way cgminer does it in OpenCL (with various lookup gaps).  In the end, the best use of the prototype hardware was actually to mine the heck out of Yacoin, since scrypt+chacha20/8+keccak(N,1,1) with N=32 (as it currently is for Yacoin until tomorrow) was almost trivial to optimize for FPGA's:

https://bitcointalksearch.org/topic/m.2127307

For the probable FPGA vs. GPU cost relationship to change, someone needs to find and disclose a method of further shortcutting the already-known TMTO of scrypt(1024,1,1).  And at that point, whatever that optimization happens to be, is very likely going to be equally applicable to OpenCL to speed up GPU processing of scrypt too.

What I have yet to see is either the BlockBurner team or jasinlee say "Oh yeah, we're genius cryptanalysts and found ways that scrypt(1024,1,1) can be calculated much faster and/or with significantly less logic than anyone else has figured out how to."

This is what worries me, and it matches a couple of other people's analyses -- people who I believe are very good at this kind of stuff.  However, if it can be done for the same price but much lower power draw, that could be a big help for those of us who have issues based on power availability.

Hardware's not my thing at the moment (it's something I want to learn in the future, but probably not for a couple of years).
legendary
Activity: 1610
Merit: 1000
Crackpot Idealist
With enough whiskey, we can solve anything!
hero member
Activity: 742
Merit: 500
Its as easy as 0, 1, 1, 2, 3
"Oh yeah, we're genius cryptanalysts and found ways that scrypt(1024,1,1) can be calculated much faster and/or with significantly less logic than anyone else has figured out how to."

Oh yeah, we're genius cryptanalysts and found ways that scrypt(1024,1,1) can be calculated much faster and/or with significantly less logic than anyone else has figured out how to.  Tongue Just kidding.

legendary
Activity: 1610
Merit: 1000
Crackpot Idealist
That is a very well thought out and technical rebuttal to my pipe dream. Thanks for ruining christmas.

My (unskilled) thinking was that scrypt is ram heavy, ram is pretty cheap these days. Math, on the other hand, is not my strong suit. Thanks for the perspective and publically abusing limitless in that thread you linked to. Wink
sr. member
Activity: 347
Merit: 250
Should I jump off a cliff covered in green jello?

Yes, but just for fun the entertainment of Youtube viewers everywhere.

Fixed that for you!
full member
Activity: 153
Merit: 100
...
Should I jump off a cliff covered in green jello?

Yes, but just for fun
sr. member
Activity: 347
Merit: 250
1. Do you think the market and community is ready for FPGA Litecoin?
Market, maybe. Community, YES! If you can pull better numbers per joule and come in cheaper on hardware cost vs gpu, people will eat em up!

This part (cheaper than GPU's for a given hash rate) is unlikely.  When it comes to scrypt(1024,1,1), a modern Radeon GPU (prior to the 7xxx series, anyway) very nearly is an ASIC already optimized for scrypt, and designed at optimal process node and with economies of scale of being a consumer product.  I've already given it a go with the Xilinx Artix-7 FPGA's surrounded by large quantities of DDR3 and tried a variety of approaches across the TMTO spectrum for the known methods of calculating scrypt, ranging from no use of external memory at all and pipelining the entire calculation, to replicating the way cgminer does it in OpenCL (with various lookup gaps).  In the end, the best use of the prototype hardware was actually to mine the heck out of Yacoin, since scrypt+chacha20/8+keccak(N,1,1) with N=32 (as it currently is for Yacoin until tomorrow) was almost trivial to optimize for FPGA's:

https://bitcointalksearch.org/topic/m.2127307

For the probable FPGA vs. GPU cost relationship to change, someone needs to find and disclose a method of further shortcutting the already-known TMTO of scrypt(1024,1,1).  And at that point, whatever that optimization happens to be, is very likely going to be equally applicable to OpenCL to speed up GPU processing of scrypt too.

What I have yet to see is either the BlockBurner team or jasinlee say "Oh yeah, we're genius cryptanalysts and found ways that scrypt(1024,1,1) can be calculated much faster and/or with significantly less logic than anyone else has figured out how to."
legendary
Activity: 1610
Merit: 1000
Crackpot Idealist
1. Do you think the market and community is ready for FPGA Litecoin?
Market, maybe. Community, YES! If you can pull better numbers per joule and come in cheaper on hardware cost vs gpu, people will eat em up!


2. Is there definite interest in FPGA Litecoin machines? Would you buy one if the price was reasonable? What is reasonable?
Yes!
Yes!
Free! That will depend on the output. What suckered me into pre-ordering a jala was the price point provide a LOT of hash for the money... If possible, make em scale. There are a ton of people who want to get in on crypto currency because of ideals or greed and typically it's easier to get your feet wet on any endevour if it does not require a second mortgage to do so.

3. Would you pre-order one to support first round funding for prototyping and first wave production?
Thats a hard one to call. I think that would tie in directly to price. Same with answer 2, it was easier to take a chance on a BFL pre-order because it was not a lot of money. Will I shell out $15k? Not on your life. Will I shell out $200 to be an early adopter? Probably.

Some thoughts: I LOVE the mere mention of escrow for funds on a pre-order on your site. If you just manage to do the opposite of everything BFL has done with their ASIC offerings, you will do very well for yourself.

I wish you luck
sr. member
Activity: 406
Merit: 250
Really hope Operatr could give us a head start over Jasinlee fpga... I hope we are still leading  Smiley
member
Activity: 67
Merit: 10
I feel as though LTC should be the currency of choice. BTC owners will have no trouble transferring their funds to LTCs.
hero member
Activity: 798
Merit: 1000
www.DonateMedia.org
In reference to pre-orders, what kind of currency are you considering as payment? Or maybe you guys haven't thought that far ahead. :x

I am not settled on that detail at the moment, but likely USD and LTC for sure, maybe BTC if I see a demand for it, but I would like to support Scrypt coins as this is a Scrypt device.



How would you all prefer to pay is perhaps a better question Smiley

Ill add this to our FAQ
member
Activity: 67
Merit: 10
In reference to pre-orders, what kind of currency are you considering as payment? Or maybe you guys haven't thought that far ahead. :x
hero member
Activity: 798
Merit: 1000
www.DonateMedia.org
I intend to keep this thread updated, certainly.

The problem now is the Altcoin board here has exploded so this thread is getting insta-buried in it, and our own forums was a gimme anyway  Smiley

Operatr
full member
Activity: 193
Merit: 100
Well, since the Altcoin forum here has quickly outgrown this place, I have set up our own  Cool

http://blockburner.net/forum/index.php






PM'd about admin capabilities and such Cheesy
sr. member
Activity: 406
Merit: 250
full member
Activity: 210
Merit: 100
Well, since the Altcoin forum here has quickly outgrown this place, I have set up our own  Cool

http://blockburner.net/forum/index.php




You should continue to update this one... 
hero member
Activity: 798
Merit: 1000
www.DonateMedia.org
Well, since the Altcoin forum here has quickly outgrown this place, I have set up our own  Cool

http://blockburner.net/forum/index.php



Pages:
Jump to: