DIY FPGA Mining rig for any algorithm with fast ROI

[GPUHoarder]

What happens if coins move to ethash-like algorithms where the efficiency gains would be alot smaller?

That’s the nature of crypto, you have to decide what risks and opportunities are right for you. Currently what I’m proposing plays well in the acceleration role with GPUs for those algorithms. In 9-12 months both FPGAs and GPUs will be shipping with similar memory bandwidth and then FPGAs (imho) will be winning on power.

Next gen Radeon Vega is rumored to have up to 1.28 Terabyte/s mem bandwith. IS there any FPGA (or announced) that would pull the same kind of perf ?

The ultraram on the FPGA collectively have a total memory bandwidth of 5.1TB/s. The HBM2 on the next generation FPGA will have the same bandwidth as the vegas with HBM2 and those chips will still have ultraram on top of that.

LOL VCU1525 has only 33 MB of uram - it is nothing. This price overvalued FPGA's can mine only few algos. If the anti-asic and anti-fpga policy will be continued, this FPGA will mine nothing.
Main problem is small amount of high speed RAM and small amount of LUT. This is kicking out a lot of actual algos like equihash, ethash, CN-v7, X17, X16r, Lyra2Z, leaving fpgas with primitive low-profit algos like keccak, tribus, skunkhash, lyra2v2, etc.

33MB is a lot for on chip memory. VCU1525 also have up to 64GB of DDR4.

DDR4 is slow. You cannot use it for ethash or CN-7

You’ve made two very aggressive anti-FPGA posts. Can you share your credentials and experience leading you to make such statements, or demonstrate why these things are not possible? You’ve made at least 4 claims that I personally know to be untrue.

[EnglishGentleman]

Anyone trying to organize a group buy of the VCU1525 kits ?

Just buy direct from avnet or the distributor. Going to be the cheapest as they are subsiding it.

Got 12 days to get your avnet order in before they increase the price. Then the group buy i'm trying to put together will be the cheapest option.

I am interested in the group buy.  How can one participate?

[jamesgalb]

So.  What programming languages and information should I be learning so that I can better understand this stuff?  In school now.

[GPUHoarder]

Hi,

Really interesting thread! I am beginning to mine using FGPA boards. I have been GPU and asic mining for the past year or so and have been interested in the FGPA side of things since hearing of it recently. I have purchased a Terasic DE0-Nano-SoC Development and Education Board to begin my learning. Can you say if any of your algo's will be compatible with the board mentioned? 1GB DDR3 ram and ARM dual processor.

Best,

Karl.

The DE0-Nano-SoC is a great board to get into learning this stuff. I have this and a few of the Max10 50kLE boards that I occasionally use for initial RTL hardware evaluation because their just so easy to work with. Vivado and Xilinx’s tools are steeper learning curve, but definitely doable.

With that said, the Cyclone V 5CSEMA4U23C6N on that board is about 1/5th the logic of what I’m proposing and while it has high-speed DMA to the ARM cores, it doesn’t have any high speed IO such as PCIe to other peripherals. That makes it tough to use outside of single small algorithms (I.e. Keccak).  The $199 version of the M.2 has ~100 Logic Elements and the $329 version has more than 200, capable of running at higher speed. The top package also has 1GB DDR3. Peak power consumption is around 15W from on-board M.2


Here’s an update summary for those catching up on this thread:

1. A few of us have been working on algorithms on FPGAs, and they’re profitable. Whitefire990 intents to release his with miner/dev fee and others of us have decided to share with the community in various forms as well.

2. On the high end the VCU1525, based on the VU9P is a very good candidate for this. It is currently a Xilinx development board on promotion for $3995, but only very small batches are being produced and the price is set to go up. Some algorithms require connecting multiple with high speed links to achieve the best performance.

3. Senseless (and possibly self, if I can be helpful) have been organizing essentially a group buy - but with FPGAs it’s less of a group buy and more of a group build. That is expected to ramp up the VCU1525 style (some power/cooling improvements)  availblility in a similar price envelope to Xilinx dev version. Working on production in US and Europe so it can be available everywhere.

4. I’ve also decided to reveal one of the smaller FPGA options in the $200-350 price envelope that I had developed for internal use and deploy, to provide an entry level option. It has a slightly different set of capabilities from the big VCU1525, but for many things it does scale.

For both hardware offerings from the community (and possibly others) orders are expected to start in June.

These boards are not so much for Ethash (though they can be used to assist/accelerate GPUs in it) , or Equihash. They excel , can improve total system performance on, or are at an advantage on Keccak (and most SHA3 candidate) , Phi variants, NIST5, Timetravel10, Lyra/LyraRev2, etc.

Someone else can chime in if I missed anything.

Damn. Watching your stuff with interest. Got any specs on the smaller FPGA option?

Which specs are you looking for? Device specs or specific algo performance?

[krnlx]

You’ve made two very aggressive anti-FPGA posts. Can you share your credentials and experience leading you to make such statements, or demonstrate why these things are not possible? You’ve made at least 4 claims that I personally know to be untrue.

Not agressive, pessimistic.  FPGA cards like that exists from 2014 https://www.alpha-data.com/dcp/products.php?product=adm-pcie-ku3
Many people(in Russia) pm'ed me and ask about this. They read this topic and think that it will be a GPU replacement.

made at least 4 claims that I personally know to be untrue.

lyra2z will not fit fully pipelined (corrected). All other algos in my post _can_ be fitted and managed to run on fpga, but they will be to slow to ROI $4000 ($5000 soon) board.

[GPUHoarder]

You’ve made two very aggressive anti-FPGA posts. Can you share your credentials and experience leading you to make such statements, or demonstrate why these things are not possible? You’ve made at least 4 claims that I personally know to be untrue.

Not agressive, pessimistic.  FPGA cards like that exists from 2014 https://www.alpha-data.com/dcp/products.php?product=adm-pcie-ku3
Many people(in Russia) pm'ed me and ask about this. They read this topic and think that it will be a GPU replacement.

made at least 4 claims that I personally know to be untrue.

lyra2z will not fit fully pipelined (corrected). All other algos in my post _can_ be fitted and managed to run on fpga, but they will be to slow to ROI $4000 ($5000 soon) board.

Anyone that thinks that FPGAs are going to destroy GPUs 10x is wrong, but being 10-200% more power and cost efficient total ROI - yes, that’s going to happen. I had a good conversation today I wish I could share, but sufficient to say that there are forces actively working to make sure the above remains true, and FPGAs have plenty of margin to play with to make that happen and continue to happen - GPUs have no room to move.

There are also some major developments over the last 4 years. The 16nm vs 20nn fabric is a big changer, and if you look at the old spartix 6 night and day. SERDES speed is light years different now, to the point that the amount of logic that can fit on a single chip is absolutely not an issue. For every chip to chip pair you can handle 50+ internal bus bits at full fabric speed.  Hard resources and URAM and even BRAM have exploded. FPGAs have gotten better, cheaper, and faster much more than GPUs have.

The last open piece is high bandwidth access to big memory. It does exist now, but you have to architect it into your system. Within two years it will be there and there will no longer be a difference for that part.

I also don’t think FPGAs are a centralization risk like ASICs, and as long as they remain closely competitive third hardware option I don’t think there will be a reason for teams to explicitly resist them. I haven’t gained a full sense of where the community stands in that however.

Edit: are you trying to fit the whole chain in static logic fully pipelined? Lots of efficiency trade offs you make to do that, and it is not strictly necessary. Chip to chip FPGAs or look at partial reconfiguration options. Gains are possible.

[R0land]

So.  What programming languages and information should I be learning so that I can better understand this stuff?  In school now.

VHDL and logic algebra

[GPUHoarder]

So.  What programming languages and information should I be learning so that I can better understand this stuff?  In school now.

VHDL and logic algebra

If you’re in the US learn Verilog - Europe learn VHDL. Or learn both, Logic is more important.

@r0land you once mentioned 64GH Cryptonight(v7?) . Care to comment on that further? What exact hardware?

[gameboy366]

So.  What programming languages and information should I be learning so that I can better understand this stuff?  In school now.

VHDL and logic algebra

If you’re in the US learn Verilog - Europe learn VHDL. Or learn both, Logic is more important.

@r0land you once mentioned 64GH Cryptonight(v7?) . Care to comment on that further? What exact hardware?

And what for asian countries ? Why does location matters here ?

[Amph]

Hi,

Really interesting thread! I am beginning to mine using FGPA boards. I have been GPU and asic mining for the past year or so and have been interested in the FGPA side of things since hearing of it recently. I have purchased a Terasic DE0-Nano-SoC Development and Education Board to begin my learning. Can you say if any of your algo's will be compatible with the board mentioned? 1GB DDR3 ram and ARM dual processor.

Best,

Karl.

The DE0-Nano-SoC is a great board to get into learning this stuff. I have this and a few of the Max10 50kLE boards that I occasionally use for initial RTL hardware evaluation because their just so easy to work with. Vivado and Xilinx’s tools are steeper learning curve, but definitely doable.

With that said, the Cyclone V 5CSEMA4U23C6N on that board is about 1/5th the logic of what I’m proposing and while it has high-speed DMA to the ARM cores, it doesn’t have any high speed IO such as PCIe to other peripherals. That makes it tough to use outside of single small algorithms (I.e. Keccak).  The $199 version of the M.2 has ~100 Logic Elements and the $329 version has more than 200, capable of running at higher speed. The top package also has 1GB DDR3. Peak power consumption is around 15W from on-board M.2


Here’s an update summary for those catching up on this thread:

1. A few of us have been working on algorithms on FPGAs, and they’re profitable. Whitefire990 intents to release his with miner/dev fee and others of us have decided to share with the community in various forms as well.

2. On the high end the VCU1525, based on the VU9P is a very good candidate for this. It is currently a Xilinx development board on promotion for $3995, but only very small batches are being produced and the price is set to go up. Some algorithms require connecting multiple with high speed links to achieve the best performance.

3. Senseless (and possibly self, if I can be helpful) have been organizing essentially a group buy - but with FPGAs it’s less of a group buy and more of a group build. That is expected to ramp up the VCU1525 style (some power/cooling improvements)  availblility in a similar price envelope to Xilinx dev version. Working on production in US and Europe so it can be available everywhere.

4. I’ve also decided to reveal one of the smaller FPGA options in the $200-350 price envelope that I had developed for internal use and deploy, to provide an entry level option. It has a slightly different set of capabilities from the big VCU1525, but for many things it does scale.

For both hardware offerings from the community (and possibly others) orders are expected to start in June.

These boards are not so much for Ethash (though they can be used to assist/accelerate GPUs in it) , or Equihash. They excel , can improve total system performance on, or are at an advantage on Keccak (and most SHA3 candidate) , Phi variants, NIST5, Timetravel10, Lyra/LyraRev2, etc.

Someone else can chime in if I missed anything.

Damn. Watching your stuff with interest. Got any specs on the smaller FPGA option?

Which specs are you looking for? Device specs or specific algo performance?

i'm also interested in the small one, if it's more fast than a 1080 or near that perf, but you said that it scale so it should be around 12 times slower or around that, this mean 2.5 faster tahn  a 1070 right?

[oskarwu]

DDR4 is slow. You cannot use it for ethash or CN-7

There is special trick possibility for FPGA to use DDR3/4 while reach the performance level as GPU in Ethash ...

Oskar

[senseless]

lyra2z will not fit fully pipelined (corrected).

Who said that it would? Who implied that it would be a good idea to fully pipeline it?

I'm pretty sure you're just fishing.

[krnlx]

lyra2z will not fit fully pipelined (corrected).

Who said that it would? Who implied that it would be a good idea to fully pipeline it?

I'm pretty sure you're just fishing.

Forget lyra2z, mtp(new algo) soon.

This thread has 44 pages now and you(fpga devs) not provided any proof, just theory.
You(fpga devs) say : go, buy fpga and later we maybe implement and maybe release some bitstreams.

I ask again : What are you going to mine with this fpga?

Let's sum up this discussion

Keccak - will work, low profit
Tribus - will work, low profit due to low network hashrate
Phi1612 - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
Skunkhash - will work, low profit due to low network hashrate
x17 - will work with 2 FPGA, need a verilog masochist, profit will be destroyed with 450x fpga
x16r - most likely will not work, or skip hard blocks at hashing,  need a verilog masochist
x16s - will work with 2 FPGA, need a verilog masochist
lyra2z - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
lyra2v2 - will work, low profit, asic soon
lyra2 - will work, low profit, low network hashrate
ethash - will work with ddr4 ram, low profit
equihash  - will work with ddr4 ram, low profit
cn-7, cn-light, etc - will work, low profit
Xevan - will work with 4 FPGA, need a verilog masochist, low profit due to low network hashrate
bitcore - I think, it will fit, and work with good hashrate(600MH), but profit will be destroyed due to low network hashrate
nist5, neoscrypt - I did not investigate, possible candidates for fpga. |

What algorithms have I forgotten?

[yrk1957]

Forget lyra2z, mtp(new algo) soon.

This thread has 44 pages now and you(fpga devs) not provided any proof, just theory.
You(fpga devs) say : go, buy fpga and later we maybe implement and maybe release some bitstreams.

I ask again : What are you going to mine with this fpga?

Let's sum up this discussion

Keccak - will work, low profit
Tribus - will work, low profit due to low network hashrate
Phi1612 - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
Skunkhash - will work, low profit due to low network hashrate
x17 - will work with 2 FPGA, need a verilog masochist, profit will be destroyed with 450x fpga
x16r - most likely will not work, or skip hard blocks at hashing,  need a verilog masochist
x16s - will work with 2 FPGA, need a verilog masochist
lyra2z - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
lyra2v2 - will work, low profit, asic soon
lyra2 - will work, low profit, low network hashrate
ethash - will work with ddr4 ram, low profit
equihash  - will work with ddr4 ram, low profit
cn-7, cn-light, etc - will work, low profit
Xevan - will work with 4 FPGA, need a verilog masochist, low profit due to low network hashrate
bitcore - I think, it will fit, and work with good hashrate(600MH), but profit will be destroyed due to low network hashrate
nist5, neoscrypt - I did not investigate, possible candidates for fpga. |

What algorithms have I forgotten?

While this may be true in the short/medium term, but in the long term FPGA may end replacing GPUs, as the costs go down and other players comes in and we see more mining oriented FPGAs. This is just the genesis of public FPGA mining.

Exceptions would be ASIC alog coins or coins who's Devs want to be FPGA-resistant (at least they think they can be FPGA-resistant).

[LTCMAXMYR]

lyra2z will not fit fully pipelined (corrected).

Who said that it would? Who implied that it would be a good idea to fully pipeline it?

I'm pretty sure you're just fishing.

Forget lyra2z, mtp(new algo) soon.

This thread has 44 pages now and you(fpga devs) not provided any proof, just theory.
You(fpga devs) say : go, buy fpga and later we maybe implement and maybe release some bitstreams.

I ask again : What are you going to mine with this fpga?

Let's sum up this discussion

Keccak - will work, low profit
Tribus - will work, low profit due to low network hashrate
Phi1612 - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
Skunkhash - will work, low profit due to low network hashrate
x17 - will work with 2 FPGA, need a verilog masochist, profit will be destroyed with 450x fpga
x16r - most likely will not work, or skip hard blocks at hashing,  need a verilog masochist
x16s - will work with 2 FPGA, need a verilog masochist
lyra2z - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
lyra2v2 - will work, low profit, asic soon
lyra2 - will work, low profit, low network hashrate
ethash - will work with ddr4 ram, low profit
equihash  - will work with ddr4 ram, low profit
cn-7, cn-light, etc - will work, low profit
Xevan - will work with 4 FPGA, need a verilog masochist, low profit due to low network hashrate
bitcore - I think, it will fit, and work with good hashrate(600MH), but profit will be destroyed due to low network hashrate
nist5, neoscrypt - I did not investigate, possible candidates for fpga. |

What algorithms have I forgotten?

agree,there was a ASIC algo defeated by GPU ,bytom ,let's see if the FPGA devs can design it.

[krnlx]

While this may be true in the short/medium term, but in the long term FPGA may end replacing GPUs, as the costs go down and other players comes in and we see more mining oriented FPGAs. This is just the genesis of public FPGA mining.

Exceptions would be ASIC alog coins or coins who's Devs want to be FPGA-resistant (at least they think they can be FPGA-resistant).

There are 3 threads, where fpga devs are selling(or incite to buy) this devices today.

Most of those algos were FPGA and ASIC resistant in past(like scrypt, x11,x13, quark, etc). But coin devs trying to make fpga-resistant and asic resistant algos. They will increase memory size and bandwidth requirements and make new resistant(today resistant, not tomorrow) algos, like mtp, equihash(with different parameters), phi2, etc.

[senseless]

agree,there was a ASIC algo defeated by GPU ,bytom ,let's see if the FPGA devs can design it.

Ya, that was pretty amazing. If their TPU is really that bad, the FPGA could probably do better than both of them.

[senseless]

While this may be true in the short/medium term, but in the long term FPGA may end replacing GPUs, as the costs go down and other players comes in and we see more mining oriented FPGAs. This is just the genesis of public FPGA mining.

Exceptions would be ASIC alog coins or coins who's Devs want to be FPGA-resistant (at least they think they can be FPGA-resistant).

There are 3 threads, where fpga devs are selling(or incite to buy) this devices today.

Most of those algos were FPGA and ASIC resistant in past(like scrypt, x11,x13, quark, etc). But coin devs trying to make fpga-resistant and asic resistant algos. They will increase memory size and bandwidth requirements and make new resistant(today resistant, not tomorrow) algos, like mtp, equihash(with different parameters), phi2, etc.

So, we wait for the next chip and launch then -- Once the stratix and virtex devices with HBM2 hit the market, there's nothing that the coin devs will be able to do. Literally, nothing.

[GPUHoarder]

So.  What programming languages and information should I be learning so that I can better understand this stuff?  In school now.

VHDL and logic algebra

If you’re in the US learn Verilog - Europe learn VHDL. Or learn both, Logic is more important.

@r0land you once mentioned 64GH Cryptonight(v7?) . Care to comment on that further? What exact hardware?

And what for asian countries ? Why does location matters here ?

Location doesn’t really matter, the US just happens to have more RTL jobs in Verilog and Europe tends to have more in VHDL, so it makes sense to learn the skill that will be most useful for you long run.

As to all the algo/profit list - some you’ve listed red are actually green. With that said, anyone who wants to buy a board to just plug in an launch a miner isn’t the target for this thread. Those who want to be on the bleeding edge and get involved while making return and profit on their hardware are best suited.

FPGA resistance and ASIC resistance are also not the same thing, both philosophically and in practice.

Finally, at least personally I have zero interest in inciting people to go out and buy random or specific FPGAs right now sight unseen. The device I’ve personally got I developed for myself, and anyone looking at chip and component costs will see that it’s true when I say I’m making it available near cost. I think I’ve explained clearly what it is designed to do, and I’m not asking anyone to buy it sight unseen. I’ve specifically declined to provide performance stats until they’re verified on the final shipping hardware and not design verification boards.

The VCU1525 is a Xilinx board that whitefire990 did his development work for. Xilinx sells that board themselves - no one here profits from inciting people to buy that board.

This is crypto - everyone needs to make their own calculations and risk decisions. I’d simply hate to see you dissuade someone from something they are interested in based on arm chair conjecture. The hardware world moves slowly - the only thing needed is patience for your proof.

[senseless]

lyra2z will not fit fully pipelined (corrected).

Who said that it would? Who implied that it would be a good idea to fully pipeline it?

I'm pretty sure you're just fishing.

Forget lyra2z, mtp(new algo) soon.

This thread has 44 pages now and you(fpga devs) not provided any proof, just theory.
You(fpga devs) say : go, buy fpga and later we maybe implement and maybe release some bitstreams.

I ask again : What are you going to mine with this fpga?

Let's sum up this discussion

Keccak - will work, low profit
Tribus - will work, low profit due to low network hashrate
Phi1612 - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
Skunkhash - will work, low profit due to low network hashrate
x17 - will work with 2 FPGA, need a verilog masochist, profit will be destroyed with 450x fpga
x16r - most likely will not work, or skip hard blocks at hashing,  need a verilog masochist
x16s - will work with 2 FPGA, need a verilog masochist
lyra2z - will work, forget it, devs changing algo due to anti-fpga and anti-asic policy
lyra2v2 - will work, low profit, asic soon
lyra2 - will work, low profit, low network hashrate
ethash - will work with ddr4 ram, low profit
equihash  - will work with ddr4 ram, low profit
cn-7, cn-light, etc - will work, low profit
Xevan - will work with 4 FPGA, need a verilog masochist, low profit due to low network hashrate
bitcore - I think, it will fit, and work with good hashrate(600MH), but profit will be destroyed due to low network hashrate
nist5, neoscrypt - I did not investigate, possible candidates for fpga. |

What algorithms have I forgotten?

I did provide proof. Screen shots, pictures, btc addresses, etc.

All any of us can say, is what we have been saying... Which is that, these devices are far superior in performance to anything you can get in a gpu. The only way to defeat them currently is with more memory. As soon as the 31-37P comes out, GPUs have lost the race in finality. These devs who are intending to stay with PoW will either need to work with us or they'll just keep getting hit by secret asics and secret fpgas.

I haven't said anything here that I didn't already say in this monero thread back in October when they were already being mined by bitmain and others.