Custom FPGA Board for Sale!

[mb300sd]

I have little experience with FPGAs, but I know that you can substitute the programmer for a microcontroller in order to load code on them, I have a ProxMark3 (RFID hacking tool, and only device I own where I can access a FPGA) that I've been dabbling with FPGA code on, it uses its ARM processor to load the FPGA. Might I suggest adding a basic one, like a PIC18 in order to avoid having to buy the expensive cable and make it easier to chain them? The PIC18F2550 has USB and can be chained easily enough using I2C. The 67J60 has a built in ethernet controller and that combined with a FPGA miner could be completely standalone.


Also, is there any chance that the FPGA miner code will run on a Spartan XC2S30? If so, I can probably create a version that will run on a ProxMark (~$300)

[vx609e]

I'll try daisy chaining them on the next board I produce. I'm sure within the Xilinx ISE it'll work fine, but outside of that (the mining script etc.) I don't think it's set up just yet to use 2 FPGAs, pretty sure some code changes in the mining script at least should take care of that.

I'm sure people will want to buy more than one board...so some daisy-chaining mecanism would be good for your business

[echo2]

wait what else can fpga board do??
i dont have much expirence with them.

diid u get your chips from http://avnetexpress.avnet.com?

[fizzisist]

Consider me in. I think the perfect FPGA miner is something in between this design and the sophisticated one we've been developing in the "modular" thread. Let's work together and get things moving, at a reasonable cost and in reasonable time.

[makomk]

Out of curiosity how much LUTs do the 32bit rotations consume?

Errrm... none, because the rotations are just rearranging bits in a fixed way. They do use up routing resources though. (I'm not involved with the design and sale of this board but I did contribute a few performance-related improvements to the FPGA code.)

[ElectricMucus]

Out of curiosity how much LUTs do the 32bit rotations consume?

Errrm... none, because the rotations are just rearranging bits in a fixed way. They do use up routing resources though. (I'm not involved with the design and sale of this board but I did contribute a few performance-related improvements to the FPGA code.)

Are you certain? Aren't the LUTs involved in routing too? I always thought that they were either configured as logic or routing node.

I have to admit my understandings of the workings of fpgas isn't quite as good.

[allinvain]

I would so buy one or two but $440 is a bit too much to invest in 100 Mhash. It would take a seriously long time to earn back that original investment. I just picked up a 5870 for $180 and that nets me 393 Mhash/s , so 3x the amount this board does at a more reasonable cost.

According to my calculations it would take  23.66 months to pay for this board - assuming btc prices stay at ~10.98.

[newMeat1]

I would so buy one or two but $440 is a bit too much to invest in 100 Mhash. It would take a seriously long time to earn back that original investment.

That's OK. People like artForz get it. What happens when only half as many bitcoins are created with every block? Or farther down the road, when miners are supported mostly by transaction fees?

What I'm getting at is-- the farther along we go, the more important computational efficiency becomes. At some point GPU's aren't gonna cut it.

*Sorry to be curt in my original reply

[ElectricMucus]

artForz

I've read the irc log and just discovered that there is a wiki entry. But are those claims reasonable?
Besides the rumors, I don't think that a asic would be that desirable even by someone with alot of budget. The feature size is much too large for these low production units and really cornering the market wouldn't be possible considering that the latest fpgas, even with all its disadvantages have the smallest feature size of any product now.
If this trend continues it might even surpass the efficiency of a low production asic or any asic for that matter if they become the dominant technology.

But the strongest reason imo is that you can't update asics, so I think he can suck it  

[makomk]

Are you certain? Aren't the LUTs involved in routing too? I always thought that they were either configured as logic or routing node.

I have to admit my understandings of the workings of fpgas isn't quite as good.

LUTs are sometimes involved in certain specialised kinds of routing, but most of the routing is done through a dedicated routing fabric in which the LUTs are embedded. As I understand it this is one of the things that distinguishes FPGAs from other forms of programmable logic like CPLDs.

I would so buy one or two but $440 is a bit too much to invest in 100 Mhash. It would take a seriously long time to earn back that original investment. I just picked up a 5870 for $180 and that nets me 393 Mhash/s , so 3x the amount this board does at a more reasonable cost.

Yeah, FPGAs very power-efficient but the upfront cost is nasty for every solution I've seen so far. On the plus side, as newMeat1 says you can keep going well after GPU miners have to give up...

Edit:

But the strongest reason imo is that you can't update asics, so I think he can suck it 

That reminds me, an interesting factoid: unless he's been deliberately misleading people, at the time he had the ASICs made he didn't actually know about the optimisation of computing H+K+W[0] in the previous round, which improves performance for free by reducing the critical path (at least on FPGAs).

[CanaryInTheMine]

Are you certain? Aren't the LUTs involved in routing too? I always thought that they were either configured as logic or routing node.

I have to admit my understandings of the workings of fpgas isn't quite as good.

LUTs are sometimes involved in certain specialised kinds of routing, but most of the routing is done through a dedicated routing fabric in which the LUTs are embedded. As I understand it this is one of the things that distinguishes FPGAs from other forms of programmable logic like CPLDs.

I would so buy one or two but $440 is a bit too much to invest in 100 Mhash. It would take a seriously long time to earn back that original investment. I just picked up a 5870 for $180 and that nets me 393 Mhash/s , so 3x the amount this board does at a more reasonable cost.

Yeah, FPGAs very power-efficient but the upfront cost is nasty for every solution I've seen so far. On the plus side, as newMeat1 says you can keep going well after GPU miners have to give up...

Edit:

But the strongest reason imo is that you can't update asics, so I think he can suck it 

That reminds me, an interesting factoid: unless he's been deliberately misleading people, at the time he had the ASICs made he didn't actually know about the optimisation of computing H+K+W[0] in the previous round, which improves performance for free by reducing the critical path (at least on FPGAs).

On optimization of mining algo (be it on GPUs, CPUs or FPGAs) is there a post that describes all available optimizations?  I'd like to learn about it.  Can someone point me to it please?  or is the computing H+K+W[0] in the previous round pretty much the only optimization so far?

[fpgaminer]

I'm mostly just the firmware designer, but I will try to chime in on a few things here for everyone's benefit

First and foremost, I do not think it has been made clear that these boards are firmware upgradeable. As improvements are made to the Open Source FPGA Bitcoin Mining project, new firmware will be generated and made available. Just like we see improvements to GPU software and you can drive more out of your dusty 5850s

Personally, I am very excited about this little board as a development platform Beats the heck out of my $1000 dev kit 


Connecting Multiple Boards
As newMeat1 has hinted at, this was being fleshed out and worked on even before these first boards were sent into production. I'll be playing around with FTDI chips and/or micros, and see if we can get a nice, simple solution that will be backwards compatible with existing boards, streamline future designs, and scale well for multiple boards.

This is really really cool and I'm glad to see a working board for sale so quickly! Not needing PCIe is like a dream and mining via usb on my old AMD duron machine would rock! Best of luck!

I have an ~8 year old laptop that could probably drive these things All current GPU mining rigs can also drive it, in addition to driving their GPUs.

Also, is there any chance that the FPGA miner code will run on a Spartan XC2S30?

The Spartan-2 series is likely too old to push many MH, if any. That particular chip only has ~1K CLBs, for example.

Are you certain? Aren't the LUTs involved in routing too? I always thought that they were either configured as logic or routing node.

Think of an FPGA as a massive breadboard, with LUTs glued onto it in columns and rows. The breadboards are your "routing fabric" and allow you to choose how you connect the LUTs, and of course you can load the LUTs with whatever configuration you want. That's your most basic FPGA architecture. Just note that there is a "routing cost" proportional to the distance you try to route. For example, connecting one LUT to another half way across your massive breadboard would lead to a massive signal delay.

On optimization of mining algo (be it on GPUs, CPUs or FPGAs) is there a post that describes all available optimizations?

This thread gives a good run down of all the optimizations being implemented on GPUs.

[CanaryInTheMine]

On optimization of mining algo (be it on GPUs, CPUs or FPGAs) is there a post that describes all available optimizations?

This thread gives a good run down of all the optimizations being implemented on GPUs.
[/quote]

thanks! i did see this before and had lost track of it. appreciate the link.  so basically similar algo optimizations were implemented on your fpga code?

[fpgaminer]

so basically similar algo optimizations were implemented on your fpga code?

Not really. Those are GPU optimizations, and are focused on reducing the total number of calculations needed to compute the final hash. That benefits the GPU's performance, but none of them would really have a useful impact on an FPGA's MH/s performance. So I haven't implemented most of them, because they aren't too helpful, would clutter the code, and my time is better spent elsewhere for now.

[kano]

I was just wondering ... why does a 6.8W FPGA need a fan?

A cheap 40mm fan is like 0.07Amps so about 0.8Watts - an increase is power of about 12% - 12% is a lot.

My 6950's I set to not run the fan when they are under 60degress.
At 350Mh/s they stay over 60degress, but a drop to something like 250-300Mh/s and they will stay under the 60degrees and not even need the fans on.
... and we are talking over 100Watts (350Mh/s is about 180Watts)

Just wondering about the decision to bother to put a fan on it all ...

[kano]

so basically similar algo optimizations were implemented on your fpga code?

Not really. Those are GPU optimizations, and are focused on reducing the total number of calculations needed to compute the final hash. That benefits the GPU's performance, but none of them would really have a useful impact on an FPGA's MH/s performance. So I haven't implemented most of them, because they aren't too helpful, would clutter the code, and my time is better spent elsewhere for now.

I presume you have done the obvious optimisations that clearly mean less calculations per double hash:
pre-calculating the first 3 rounds and not bothering to calculate the last 3.5 rounds since you don't need them at all ...

[inh]

I was just wondering ... why does a 6.8W FPGA need a fan?

A cheap 40mm fan is like 0.07Amps so about 0.8Watts - an increase is power of about 12% - 12% is a lot.

My 6950's I set to not run the fan when they are under 60degress.
At 350Mh/s they stay over 60degress, but a drop to something like 250-300Mh/s and they will stay under the 60degrees and not even need the fans on.
... and we are talking over 100Watts (350Mh/s is about 180Watts)

Just wondering about the decision to bother to put a fan on it all ...

6.8W isn't a lot compared to a GPU but it is a TON compared to these FPGAs that normally run at 3 watts or below. FPGAs are really being pushed in these applications.

[newMeat1]

That's a smart question...

You would be surprised how much heat 6W can generate. I talked about this a long time ago with fpgaminer. He said something like, "It's not so surprising that you need to put a fan on an FPGA. The surprising thing is that a 6990 doesn't burn your house down."       Maybe the difference is, a 6990 already gets some airflow, being inside the computer case.

But yeah, we might be able to get away with no fan. I think the chip doesn't start incinerating 'til over 100C. fpgaminer was working on an error-detecting code. If the chip gets too hot and starts making mistakes, it will detect the mistakes and stop.

You bitcoin crowd sure are a tough sell. I think I will try to sell these on eBay

P.S. The board has been running non-stop for about 18 hours now    

[ngzhang]

I was just wondering ... why does a 6.8W FPGA need a fan?

A cheap 40mm fan is like 0.07Amps so about 0.8Watts - an increase is power of about 12% - 12% is a lot.

My 6950's I set to not run the fan when they are under 60degress.
At 350Mh/s they stay over 60degress, but a drop to something like 250-300Mh/s and they will stay under the 60degrees and not even need the fans on.
... and we are talking over 100Watts (350Mh/s is about 180Watts)

Just wondering about the decision to bother to put a fan on it all ...

It is because the FG484 chip is about 28mmx28mm, plastic packaging. Only a 28X28mm heatsink can add to the chip. So a small fan is needed.

[newMeat1]

Here's something else I just realized. Li wrote this in another thread earlier today:

In the end with my layout I was forced to use a small heatsink (2.5x2.5cm foot print) with a small fan, couldn't find a heatsink that was super tall to get away with a passive design. The heat output itself at 100Mhash/s is significant, you can actually measure an increase in current consumption as the part heats up (possibly due to leakage increasing), and drop when u cool it.

So if the chip gets hot, it starts using more power. Looks like the ~0.5W of the fan is justified...