Bitcoin Forum

There seems to be quite some possibility to acquire fpgas on ebay.

Most of them are from damaged boards where you would have to reball them yourself, but I think we woudn't necessarily need that.

My thought is get them off the board with a hot air gun and glue several of them on a big heatsink upside down with a heat-conducting glue.
They possibly couldn't be used for anything else after that but what the hell, they are most likely living their rest of their lifespan generating bitcoins anyway...

Then solder wires directly onto the pad were the reballing would take place, that could be quite difficult but I think it would be possible.
The optimal way would be to devise a system were a minimal amount of wiring is required, only jtag + serdes nodes to get the data though. (or possibly even with the jtag backchannel, would be even more awesome)

The only real isse I see here is supplying power, since they sometimes require different levels and designing differnt power converter circuity everytime would not be very efficient and would require too much work.

Possible sources:

http://www.ebay.com/itm/Xilinx-Virtex-5-XC5VLX110T-64-x-Samsung-XCK9F8G08U0M-PCB0-8GB-Chip-Recovery-/260871631798?pt=LH_DefaultDomain_0&hash=item3cbd2933b6
(might be a good board in case anyone wants to grab it and run with it go ahead)

http://www.ebay.com/itm/Lot-100-pcs-XC2V250-5FG456I-Xilinx-PCB-BGA-FPGA-/270846090663?pt=LH_DefaultDomain_0&hash=item3f0faf5da7
100 small small old virtexes, might work without unrolling.

What we also would need is a design which can be quickly adopted to many different chips with different I/O configurations.
I think if this works out this could very well be the next big thing in bitcoin mining... ;)

fail.


looking at those pics from ebay, those chips look beat to hell.

I've try that. No can do... 1mm spacing is to small. I've used a chipset from damaged matherboard. And I'm rather experienced in soldering (11 years in work and other 10 as a hobbist).

The first board's origin is kind of interesting, it looks like it's a custom board from a company that pissed away $16 million+ in venture capital in three years. Their site is all buzzwords: "Blackwave designed and built the first purpose-built video streaming appliance. Precisely matching intelligent proprietary software with industry standard hardware, Blackwave can drive higher performance while occupying less floor space and requiring less administrative oversight and power consumption." Here's a list (http://web.archive.org/web/20100127101258/http://blackwave.tv/company/magagement.htm) of their Management Team archived before the domain expired, so you know who not to hire.

   Those first ones are almost twice the I/Os of the ones being used now. They appear to be 1136 pin verses the 484 of current design. They are lower voltage than the 484 package, being 1v instead of 1.2v.   However a LOTTTT cheaper on this ebay auction than new.

  http://avnetexpress.avnet.com/store/em/EMController/Programmable-Logic/FPGA/_/N-100235/Ne-100000?action=products&cat=1&catalogId=500201&cutTape=&inStock=&langId=-1&myCatalog=&proto=&regionalStock=&rohs=&storeId=500201&term=XC5VLX110T&topSellers=

Cheaper than new but not worth it. Propably it may achieve somwhere around 150MH/s. New Spartan6 LX150 (175$ and will work for sure) can do more.

BTW. Those XC2V250 may achive somwhere near 1MH/s...
Worth to look are Virtexes 5 and 6, preferably 110+ at prices around 100$. Best buy will be XC6V240T (400MH/s possible).

  That's what I was wondering. They are almost twice the I/O but are 65nm and lower voltage. IF they can't run the same speed as the S6 then, as you say, we would not get twice the hash. You can buy an S6 -N3 quality for $158 now.


  Edit; Those second ones are a smaller package, 456 pin.  Though, with those at $2.5 per chip. If you could even get 1/4 of them in good condition with reballing might be worth messing with.  Not sure what hash they are capable of. Someone that knows what they are reading should be able to guess from the cell and i/o count on them though.

While scrap FPGA may be a viable business you need to find higher end chips.

You can't hand solder leads to BGA.  If you don't believe me take a busted motherboard desolder a northbridge chip and try to do it by hand.  After about 5 minutes you will realize it is next to impossible.  If you had to do one because it was a life or death situaiton then maybe but doing hundreds of them all at perfect quality ... not gonna happen.

You can have them reballed but it will cost some money.  The best bet would be to find an assembly company which could reball them and attached them to a daughter board.  Then you could have a set of PCB already constructed with a socket for the daughter board.

All that has overhead so it really would only make sense if you find a large lot of scrap V5 or V6 chips.

Good rolled 128 cycle design its aproximetly 2k LUT. Those have 3k LUTs. So, somwhere around 1MH/s is possible...

   I knew one of you brainiacs would be able to calculate that in short order. =)

  As DandT pointed out, it will have to be considered the cost of having a company reballing. Which, still does not add a whole lot of costs.

 I think overall, your hunt may be fruitfull when the things he pointed out are considered. Make sure to take Dexter with you so he can quote you the estimated speed of anything and keep in mind a roughly $20 per chip reballing and you're all set for a hunt.   :)

That's bad...  :-[

But what about not soldering the pad to the wire but verse visa? What do I mean?
It is conceivable that the pad doesn't provide enough area to get a solder tip a wire and the solder to it at the same time, however:

If we first make a solder ball on the pad like we would do with normal reballing, (wouldn't have to be hq)
and coat the wire with solder also we would have a very tiny area where we can have a contact. So only the end diameter of the wire would be needed (and a very steady hand & a good loupe)

So the thing to do is put the solder top against the wire only and use it to conduct the heat to the pad via contact to the ball.  

I might try this out on some damaged phones and report...  :)


PS: about those vitexes: that was just an example, I personally wouldn't bother with such puny chips but they go upto 750mhz according to the auction... Now if we can find higher end chips like that now that would be awsome...

  I've almost 0 experience in soldering anything so small but picturing the physical properties of even high grade solder, I would think when you went to place your wire through the solder into the contact point it would pentrate the ball without grabbing any solder, or moving enough of it with the wire to make the contact.  You may get lucky and have your ball hold the wire in palce just enough to make contact, though I would not trust it to remain after a few times of current flowing through it.

Hand soldering BGA's are out of question. I've managed to solder a wire to pad but I had to give up trying to solder second one. Just not enough space. However, prototype PCB with 2 layer and 1mm spacing via's can make possible to solder those BGA's. All you need its to fill few vias and position BGA, then you can easily solder wires where needed.

750MHz are just for one gate, no chance to get even close to that with complicated design. For example Spartan 6 are rated to 400MHz, best design so far reached 195MHz.

Can you refer me to a company that reballs BGAs for $20/ea?  I screwed up a few Spartan 6's during my early prototyping runs, and if it's really that cheap I'll have them reballed.

I don't know if this company is full of shit but they claim easy hand reballing using a stencil. 

http://www.solder.net/ezreball.asp

If you only have a few that may be a better route.  I doubt an assembly house wants a job to reball a half dozen chips and if they do it likely will be more than $20.

  Aye, sadly I pulled that number out of my ass.  I would check with these guys here; http://www.etech-web.com/bga-reballing.htm  only a few of them are likely more than $20.  In large quantities, 200+ it is possible to get it that cheap.

  What DeathAndTaxes pointed out is interesting. i also noticed quite a few videos claiming to detail working reball processes. I did not view any of them however.

That stencil method still requires a reflow machine.  :-\
Also why pay 105 USD for a stencil you can make with any laser cutter?

Well I left my laser cutter in my other pants.

Convection oven or electric skillet = ghetto reflow machine.  

This is the "ghetto FPGA" thread right?

Well, renting a laser cutter is available in almost any city... But... do you have some instructions on how to build a reflow out of an oven?

I mean baking pcbs in one of those small electric ovens is everymans work but reflow ???

If you guys just need a "cheap" FPGA platform for personal development, WITH OUT "open-box-can-use" mining code, that's really an easy case.

   E-tech responded to me rather quickly on a quote request.

  Few questions;
   1. Are the chips affixed to the PCB?
   2. Pb free or SnPb?
   3. How many do you need reballed?

Toaster ovens are not very good for BGA assembly - skillets/hotplates are much much more effective.
As you can monitor the process in real time with an IR thermometer.

-rph

http://zremcom.ru/images/stories/Stat/interes/2011/kulibin/rebol.jpg

someone was bored lol

Extremly bored. But I can imagine that with very good soldering skills and tools it can be done. I've tried and I failed (maybe method was bad). Anyway I think that 2 layer PCB just with vias at 1mm spacing should make things much easier and give possibility to solder again BGA chip without reballing. Also, That PCB can serve as prototype board to new chips. Alot of soldering but changes can be made easily and quickly.

Such a "solution" is just a waste of time, as the inductance of the supply wires (core voltage, I/O voltage and GND) would be way too high and the FPGA would not operate reliably at any decent clock, if at all.

Consider a FPGA implementation of the Bitcoin mining algorithm (double SHA-256). There are 128 rounds of 256 bits each, i.e. 32768 flip-flops, all of which switch at the same time. All these flip-flops switching at the same time causes a momentary spike in power draw. Let´s say, for the sake of argument, from about 1 Amp when no flip-flops are switching to, say, a momentary power draw of 25 Amps (est.). If there is any inductance at all in the supply wires, the momentary power draw of 25 Amps will cause the internal core voltage to sag below the minimum acceptable core voltage, and the ground level that the FPGA sees to rise.

Below minimum core voltage, correct operation of the FPGA is no longer guaranteed, as the contents of flip-flops and even the contents of the configuration registers may be lost or corrupted.

Even I know that there are decoupling capacities needed to avoid that. One could solder the capacities onto the power/gnd pins and the power lines onto them. The issue you described matter mostly for signals outside the chip which we practically don't have. The pins used to get the initial hashes in and results out would need minimal bandwidth.
Granted some clever i/o protocol would be needed but that wouldn't really be an issue.

Since the appearance of this photo I might even reconsider my idea.
The thing however is I yet have to find a offer where the thing would be really worth it.