Buying FPGA on eBay

[Professor James Moriarty]

 There are lots of fpga looking things on eBay , how can I know if they are any good for mining ? What specs should I be looking for?

[Gator-hex]

It needs to say Bitcoin FPGA, if it don't it's just for an electronics project. 

[razorfishsl]

It is not worth it.

Consider that even if you pick up a V5 110 for about $50 bucks

Jobs

1. trace out circuit in an attempt to find where to tap into FPGA
2. Hope to GOD that :
A. The vias you need on the pins are big enough to be soldered
B. The important pins are not tied to gnd or to existing circuit
C. There is a clock circuit and it is accessible
D. you design Suitable USB/RS232 I/O
E. The PCB powers up.

OK you have a rewired working PCB

1. System to LOAD the FPGA
2. mining software to support the IO YOU have to design
3. Layout of existing scrap PCB is suitable for the power & frequency you need...


Hurray.... step 3

Design a switched mode PSU to handle the kit, yep the PCB *may* have SM on it, but you still need to supply that with the output of another power supply..


Step 4  hurrayyy.......

You are mining @ 100-200 Mh/s


In several months you can make your $50 bucks back.......

If you have 1.7GH/s you can make 10 bucks a day when bitcoin is at 100USD each....



I have this down to a fine art....  (I can build a  2GH/s rig in a DAY!!!)

Even buying FPGA at $30USD each.... it is barely worth it with the difficulty changing 10% twice a month.....

[tony_357]

If you want to run fast and efficiently on an FPGA, you need to maximize clock rate, and throughput rate.  This essentially means a hash every clock cycle, and you fully pipeline the design (which introduces latency, but preserves throughput) to run at the max toggle rate of the FPGA.

A fully pipelined, and unrolled SHA256 engine with routing will probably run you around 150K logic cells.

If you run at 200 MHz, then you should get about 200 MHash/second per 150K logic cells.
(This is about the specs for the various Spartan 6 LX150 boards out there).

You won't find these boards for $50.

The closest is this one (4 LX150's running at 800 MHash/sec) - for $1069.99
http://modminerquadstore.com/index.php?route=product/product&product_id=50

I'm looking at putting together a dual Cyclone V based FPGA board.  This is 28 nm technology, which means I should be able to run at 250-300 MHz, and the chips are 301K LC each, so I should be able to run at about 1000-1200 MHash/sec.  If I get it going, I'm thinking of shipping units at $900 per 1000-1200 MHash/sec.

http://www.raspberrycoins.com/

There's an FAQ I put together at : http://www.raspberrycoins.com/fpga_miner


The thing is, there is this thing called ASIC mining which may be around the corner (but it's been around the corner for several months), and if it comes into being, will run faster, at a lower cost than FPGA mining.

But in the mean time, there's bitcoins to be mined at the current difficulty level, so I want the bring the question to the general market.
If it makes sense to mine bitcoins with the latest FPGA (and thus get a speed boost), let me know by registering.  If we get enough registrations, then, we'll do a crowd source funding and make some boards (delivering boards + software to our funders).

I'm essentially putting it up to the collective wisdom of the crowd to determine if it makes sense to mine bitcoins with FPGA's.

Best Regards,

Tony

[razorfishsl]

If you want to run fast and efficiently on an FPGA, you need to maximize clock rate, and throughput rate.  This essentially means a hash every clock cycle, and you fully pipeline the design (which introduces latency, but preserves throughput) to run at the max toggle rate of the FPGA.

Statement of the obvious with some gobbled de gook thrown in to look good...... (actually you need to maximize parallelism)
SCRYPT is an example of WHY  a faster clock rate will not come up with the goods on an FPGA......

You won't find these boards for $50.

YES YOU WILL if you know where to look..........

I'm looking at putting together a dual Cyclone V based FPGA board.  This is 28 nm technology, which means I should be able to run at 250-300 MHz, and the chips are 301K LC each, so I should be able to run at about 1000-1200 MHash/sec.  If I get it going, I'm thinking of shipping units at $900 per 1000-1200 MHash/sec.

Have you even bothered to simulate it?



it is still costing you close to 1k for 1.2gh/s

To generate $7us a day!!!! (or about 50 cents an hour, less than an Indian worker)
http://tpbitcalc.appspot.com/?difficulty=10076292.8834&hashrate=1200.00&exchangerate=121.56&bitcoinsperblock=25.00&rigcost=900&powerconsumption=80.00&powercost=0.10&investmentperiod=355

Like I said I can build a STABLE ~2 gh/s (non-ramped) for $300us, which is 6gh/s for your 1gh/s but both are 3 times the cost of a similar BFL product (5GH/s@274usd)........  except mine gets closer to the hash rate!!!


Plus by the time you cloud source and do your pcb it is going to be July
Why bother?

[tony_357]

Wow - somebody sounds defensive.  

Statement of the obvious with some gobbled de gook thrown in to look good...... (actually you need to maximize parallelism)
SCRYPT is an example of WHY  a faster clock rate will not come up with the goods on an FPGA......

What does SCRYPT have to do with Bitcoins?  I thought SCRYPT is Litecoins.  Or will your rig also mine Litecoins?

You won't find these boards for $50.

YES YOU WILL if you know where to look..........


Like I said I can build a STABLE ~2 gh/s (non-ramped) for $300us, which is 6gh/s for your 1gh/s but both are 3 times the cost of a similar BFL product (5GH/s@274usd)........  except mine gets closer to the hash rate!!!

From you link it appears that you are affiliated with Butterfly labs (you have a butterfly labs logo on your site).
I fully disclosed what the expected hash rate is, and that's what I can expect from an FPGA miner.  Specifically, this topic was about an FPGA miner.  So that's what I expect I could do with state of the art 28 nm FPGA - that works out to $1.33 - $1.11 / MHash/sec.

I already said that an ASIC miner would provide better performance than a FPGA miner.  That should be obvious. The only advantage of an FPGA for mining is that when the card becomes obsolete for mining, the cards can be repurposed later for other activities - like ASIC prototyping, and FPGA simulations of ASICs, or in general acting as a general purpose high speed computation platform.  A dedicated ASIC should provide better price/performance than an FPGA.

So, I clicked on your buy now button, which took me to the Butterfly labs web site - which indicated that there is at least a 2 month delay from when the pre-order gets processed.  It appears to me that you might have some knowledge about the Butterfly Labs solution.


Do you know how many of these $1299-60GHash/sec units have shipped?
(I've been checking message boards, and I mostly see that units have been experiencing delays, and that only the 5 GHash/sec units may possibly have shipped).  If you could provide details, and links that would be great.  There have been a lot of impatient customers.

If the ASIC units aren't shipping, then it might make sense to mine as soon as possible, and we can have a state of the art FPGA rig going pretty quickly.

Oh, and your link says to send us your bitcoins.  I think one advantage of our solution is that the cost is based on dollars, so as the $/BTC goes up, the price of our FPGA based rig drops substantially.

[razorfishsl]

Not defensive , just lack of research gets my back up...., because  noobs enter the site then think they are going to get rich
by spending  several grand on kit that has ROI about the same time as the sun is due to go cold.
This does nothing for the long term goals of Bitcoin.

The SCRYPT reference was to highlight the mostly incorrect statement about Clock speed.
Since we were talking about FPGA, clock-speed and parallelism, it is not specifically related to bitcoin but rather  design.
I.E if I have  an algorithm that takes 1mS to execute, then it does not matter if my processor/logic is operating at Thz speed..

But if I can parallel that algorithm to get 1,000,000 at a time, then even though it takes 1mS I am actually getting a theoretical throughput of 1ns 'each'.



I have a BFL logo on my site?
Are you ill?, that was a web link to Butterfly labs...

All that aside...  the difficulty climbing at more than 20% a month, most sites are dropping PPS, which means that your  solution will not pay itself back any time soon..... no matter what price the BTC is at.

And yes my rigs WILL mine litecoins via FPGA but I'm still working on the memory interface ,because the pure FPGA implementation currently has a shite throughput

Ohhh. and your 'website' is down.

[dropt]

I.E if I have  an algorithm that takes 1mS to execute, then it does not matter if my processor/logic is operating at Thz speed..

How do you calculate this 1ms run time if it's not a product of clock frequency?

[Professor James Moriarty]

 That fpga deal was for people who believe asicminers wont come out anytime soon I believe . I for one hand believe they already started shipping so the ROI on that fpga is getting longer and longer everyday but a person does not believe that and believes there is at least 2-3 months before enough asicminers will come to change the difference on the difficulty of bitcoin mining so they can look for a better option that is currently available. So you are talking about two different mind set customers , you are both right on different levels.

 I wanted to check some fpga system for low price so I may get some crazy discounts like : http://www.youtube.com/watch?v=jW1hFQndX2I  this guy. Otherwise if am going to pay 1000$ish I would buy 3-4 5gh butterfly and wait for them to arrive and when they arrive I will have about 20ghash for 1000 which will have horrible but would be 20 times of 1ghash for the same price.

[razorfishsl]

LOL.....

from the site:
http://www.raspberrycoins.com/calc
what a crock of shit.......

The fantasy.......




The reality......



This is EXACTLY what I am talking about Bitcoin getting a bad reputation for scammers......

[tony_357]

...Yes my rigs WILL mine litecoins via FPGA but ...

Great - show me a link to your hardware solution, or your plan for a hardware solution, or whatever.  I might want to buy one

I showed you my link : http://www.raspberrycoins.com

Since we were talking about FPGA, clock-speed and parallelism, it is not specifically related to bitcoin but rather  design.
I.E if I have  an algorithm that takes 1mS to execute, then it does not matter if my processor/logic is operating at Thz speed..

Well, earlier I said, : "A fully pipelined, and unrolled SHA256 engine with routing will probably run you around 150K logic cells."
That's an area number (ie: resource utilization).

Then I said : "If you run at 200 MHz, then you should get about 200 MHash/second per 150K logic cells."
That's because the throughput is a result every clock cycle - based on a FULLY PIPELINED, AND UNROLLED SHA256 ENGINE.
(those in caps means that it's operating in parallel - sorry if I was unable to communicate that to you or that you were unable to understand that - perhaps you thought that was "gobbled de gook thrown in to look good")

I thought it was obvious that I am expecting (for Bitcoin Mining):
 - similar FPGA resource usage numbers (per SHA256)
 - being able to stuff two SHA256 cores into an FPGA that's twice the size (ie: 301K LC Cyclone V vs LX150)
 - an increase in clock speed due to using 28 nm technology (vs the LX150's)
 - a decrease in power consumption (compared to LX150's)

Anyways, that's what I'm planning - if there's interest, great - I'll sell some, otherwise, it's no big deal to me.

[razorfishsl]

I.E if I have  an algorithm that takes 1mS to execute, then it does not matter if my processor/logic is operating at Thz speed..

How do you calculate this 1ms run time if it's not a product of clock frequency?

Ye gods......  Ok....Ok... count to 10

Take for example the access of Drams/Srams  where the setup times and charge times are a function of the architecture NOT the overall clock frequency of the system.

I.E lets say the 'setup time' of the address is 3ns, it is still 3ns no matter how fast the central clock is going.
(unless I specially select temperature or dies and pull out chips that can manage the  'setup time' in 1.5ns)

What this means is that by designing in certain requirements into the algorithm, I can to a larger degree make the overall clock rate irrelevant.

[razorfishsl]

That fpga deal was for people who believe asicminers wont come out anytime soon I believe . I for one hand believe they already started shipping so the ROI on that fpga is getting longer and longer everyday but a person does not believe that and believes there is at least 2-3 months before enough asicminers will come to change the difference on the difficulty of bitcoin mining so they can look for a better option that is currently available. So you are talking about two different mind set customers , you are both right on different levels.

 I wanted to check some fpga system for low price so I may get some crazy discounts like : http://www.youtube.com/watch?v=jW1hFQndX2I  this guy. Otherwise if am going to pay 1000$ish I would buy 3-4 5gh butterfly and wait for them to arrive and when they arrive I will have about 20ghash for 1000 which will have horrible but would be 20 times of 1ghash for the same price.

That guy on 'youtube' is a jack-off merchant, the tech he is holding is so out of date (look at the size of the RAM chip!! & the card date 01/12/00), you could possibly...... possibly mine at a couple of Mh/s, IF you could interface the chips together in a reasonable configuration.
Unless you went 'Solar' it would use more electricity than it would pay for.

Look I can see where you are coming from.. but the heady "get rich quick" days of bitcoin are OVER.
Keep your $1,000 and wait till BTC crash down to $50-$60 each then buy some (MTgox), then sell high, you will make way more money.

Once Avalon start to ramp up production (they are already shipping and even with 900 units have totally fucked the BTC 'difficulty level' from 3 million to 11 million) and when BFL start to ship, the whole thing is going to be a shit-storm as the difficulty crashes out into oblivion.

If I can knock up an FPGA rig in a day.. then why am I not filling my  house with them?
 (I recently turned down a lot of 50 V6 Xilinx FPGA @$20 bucks a card), it would have been about 15GH/s for $1,000usd and 5 days work to commission them, power usage would have been about 1Kw

I'll tell you why... because most pools are getting rid of PPS and the other 'shares' are paying so low that it is not worth it, unless you are into  tens of GH/s
 

[tony_357]

I.E if I have  an algorithm that takes 1mS to execute, then it does not matter if my processor/logic is operating at Thz speed..

How do you calculate this 1ms run time if it's not a product of clock frequency?

Ye gods......  Ok....Ok... count to 10

Take for example the access of Drams/Srams  where the setup times and charge times are a function of the architecture NOT the overall clock frequency of the system.

I.E lets say the 'setup time' of the address is 3ns, it is still 3ns no matter how fast the central clock is going.
(unless I specially select temperature or dies and pull out chips that can manage the  'setup time' in 1.5ns)

What this means is that by designing in certain requirements into the algorithm, I can to a larger degree make the overall clock rate irrelevant.

... so, this design of yours is going OFF CHIP (off the internal high speed dual port static RAMs built into the FPGA), and storing data on external DRAM. ...

... so sure - if you do that, then your clock speed is probably irrelevant - you'll be limited by the access time to go on and off chip.

[razorfishsl]

This is why I do not post or reply to posts very often.

Go re-read the  3 way split  (WAY OFF TOPIC) thread and my 3 separate answers to each section.

You will see that this answer was related to a question where the poster was under the assumption that  clock frequency was the sole indicator of the time it takes an algorithm to run.

I merely pointed out that there are times when this does not apply, SPECIFICALLY as an example I highlighted the SCRYPT algorithm that is SPECIFICALLY designed to try and make any attempt to parallel its calculation. go outside the main silicon infrastructure for resources.

and it is THAT infrastructure that sets the overall speed.

I.E you may have a porsche that does 250Mph, but if the roads in Liverpool only allow you to travel at 30 Mph, then it is irrelevant how fast your porsche can go, and the porsche does not set the journey time... the roads do.


To answer your question, If you have to travel about in Liverpool, then you have to get out of your driveway.........

[tony_357]

This is why I do not post or reply to posts very often.

Go re-read the  3 way split  (WAY OFF TOPIC) thread and my 3 separate answers to each section.

You will see that this answer was related to a question where the poster was under the assumption that  clock frequency was the sole indicator of the time it takes an algorithm to run.

I merely pointed out that there are times when this does not apply, SPECIFICALLY as an example I highlighted the SCRYPT algorithm that is SPECIFICALLY designed to try and make any attempt to parallel its calculation. go outside the main silicon infrastructure for resources.

and it is THAT infrastructure that sets the overall speed.

I.E you may have a porsche that does 250Mph, but if the roads in Liverpool only allow you to travel at 30 Mph, then it is irrelevant how fast your porsche can go, and the porsche does not set the journey time... the roads do.


To answer your question, If you have to travel about in Liverpool, then you have to get out of your driveway.........

Sure,

So, each Cyclone V  (we're looking at using two A9 parts) has (according the Cyclone V datasheet http://www.altera.com/literature/hb/cyclone-v/cv_5v2.pdf) :
  12,200 in M10K blocks
  a Hard Memory Controller - capable of 400 MHz DDR3 operation (ie: if they did this right, it should be no-muss, no-fuss, Hard IP - but who knows - I didn't look into their hard memory controller in detail, but I would be surprised if they screwed it up).

We're putting two Cyclone V's to a board which gets it's configuration bitstream downloaded from a raspberry pi.

Right now, we're trying to get a dual Cyclone V based board out as soon as possible (so we won't hook up that hard memory controller to anything), but with the next version, we'll throw in some wires to connect up that Hard Memory Controller to external DRAM chips.  For now, we're going with just the bare FPGA metal (ie: just the M10K's built into the FPGA).

oh, we'll also allow for communication between the two chips using high speed LVDS channels

Best Regards,
Tony

http://www.raspberrycoins.com

[dropt]

Ye gods......  Ok....Ok... count to 10

Take for example the access of Drams/Srams  where the setup times and charge times are a function of the architecture NOT the overall clock frequency of the system.

I.E lets say the 'setup time' of the address is 3ns, it is still 3ns no matter how fast the central clock is going.
(unless I specially select temperature or dies and pull out chips that can manage the  'setup time' in 1.5ns)

Sure, but what you've described is latency which (unless I'm mistaken) partially determines the maximum theoretical speed achievable.  This could be a primary motivating factor for pipe-lining, but it does not necessarily determine the total execution time of an "algorithm", at least not so far as removing the dependence on clock frequency.  For each hash to be completed, or for sake of conversation, the run-time of the algorithm in generating the output includes all of the latencies inherent in the design.  This included the latencies you mention, but is also dependent on the clocks per instruction (CPI), instruction count and clock time.

Execution time (T) = CPI*Instruction count*clock time = CPI*Instruction count/frequency

No amount of fixing latency times or CPI or instruction count removes the clock time from the equation.  If by:

What this means is that by designing in certain requirements into the algorithm, I can to a larger degree make the overall clock rate irrelevant.

you mean to set the collective CPI and IC to values sufficiently large enough that small (or reasonable) variations in clock time are irrelevant then you have a lot of optimizing to do.

EDIT:  I just read your other post again, I believe you're correct in latencies of off-chip resources being a large constraint (especially for scrypt). My rhetoric was in effort to not think of external bus and device latencies as being the only point of concern, which is ironic since you were trying to address the issue of on-chip timing not being the only concern.

[ultraimports]

I'm sorry but...

The OP was simply asking what to look for on Ebay. Just because you disagree with HOW they want to mine, doesn't mean we need the entire breakdown. Personally, I see no issue mining FPGA, even at a premium.

Since I have been moving rigs around, changing out video cards, etc, I have found that some of the AMD drivers give me fits and I end up losing an hour or two simply removing and re-installing drivers for different generations of cards -- what a pain!

Lets not forget the power consumption and HEAT the GPU mining rigs put out. Also, it would be nice to pause my miners for the occasional FPS game to vent my frustrations out online....can't touch it while its mining right now...ugh...

So, if someone wants to drive 250mph in a 30mph zone, let them ...not our issue.

OFF_SOAP_BOX

[Professor James Moriarty]

Exactly , maybe there will be asic miners maybe there wont be any (at least any time soon) I believe there are not enough asic chips or miners anytime soon and I was considering mining with fpga , there are people with 30+ fpga miners that makes a real good ROI and I wanted to be one of them

[clowe]

I just purchased a Bitcoin FPGA ModMiner Quad 800MH-1GH/s miner board for $300 off Ebay, a bit of an impulse buy but if it's legit seems like a bargain.  I figure since its on eBay and I used Paypal if its a scammer I'll just claim back my money.

The same guy was also selling the Block Erupter USB miners and people were prepared to pay $331 for them, nuts since they only mine at 300MH