DIY FPGA Mining rig for any algorithm with fast ROI

[s1gs3gv]

Any comments on the quality/portability of the athena code ?

[1715389171]

1715389171

[1715389171]

1715389171

[1715389171]

1715389171

[1715389171]

1715389171

[1715389171]

1715389171

[mrb]

It would be nice to have a complete database documenting all PoW algos (including pseudo-code). If someone feels motivated to do it, I'd be glad to sponsor them.

Can you please elaborate on the sponsorship. I understand the equihash algorithm very well, by the way.

It would be very informal. Start writing the documentation for 3-4 PoWs. If it's high quality and think you have what it takes to continue, I'll periodically donate you some BTC whenever you add more PoWs to the doc ($10-20 per hour of work).

[HardwareCollector]

It would be nice to have a complete database documenting all PoW algos (including pseudo-code). If someone feels motivated to do it, I'd be glad to sponsor them.

Can you please elaborate on the sponsorship. I understand the equihash algorithm very well, by the way.

It would be very informal. Start writing the documentation for 3-4 PoWs. If it's high quality and think you have what it takes to continue, I'll periodically donate you some BTC whenever you add more PoWs to the doc ($10-20 per hour of work).

I can find a college student or two who may jump on that opportunity, we’ll see.

[HardwareCollector]

Let’s play the devil’s advocate.

While all of these sounds good, there are some risks that you can’t afford to ignore:

Each hardware that is brought online, decreases profitability for all other existing hardware (based on hash rate).
Everyone is fighting for a piece of the same pie.

The early adopter has the shortest time to ROI before the coin becomes saturated with hash rate.
Things always look good if you ignore the eventual increase in global hash rate.

You are at the developer’s mercy for releasing FPGA bitstreams if you can’t design your own.
Variations in performance can vary widely between public and private designs.
You are locked to a few small coins and someone with a better algo implementation can decimate your ROI time.

And the biggest risk of them all is AWS F1 instances.
As long it’s profitable to mine on AWS or other cloud providers, people will do so. Amazon does not shut down your instances when you are upfront with them, and how do I know this? They care about your ability to pay bills and follow their ToS.

Just playing the devil's advocate.

this is a major concern. Just how long before my 5x 5000 = 25000 fpga get the software upgrades I need.

There was some safety in a world of 200,000 gpus on a coin  since many wanted some software and a developer Claymore for instance could simply take 2% off the top.

In this world of fgpa there will not be 200,000 pga's screaming please give me new software upgrades.

and as I said  if the coin  changes algos a lot  even the pga guy will struggle to keep up with the switches.

and if the coin never changes the asic builder will whale.

so the fpga guy will need just enough algo changes to be okay.

I suspect
 the fpga will be good for some coins
the asic guy will be good for some coins
the gpu will still do okay.

If you are on the fence about purchasing, then it is always best to wait and see how things play out. Most, if not all, industrial miners already understand the economics of mining quite well, and there’s no need repeating it here. But what some people do not understand is that these FPGAs are good for saving space and power, which can be appealing to some users.

Now let’s see what happens when the Keccak algorithm (assuming 17 GH/s) is released on May 30th and you are able to mine with it. Hmmm, your profitability will be on par with other GPU miners on other coins with the same overall costs. You haven’t lost anything, but at least you now know the OP is credible and have some space and power savings. Then guess what happens next, there’s going to be a massive influx of miners who will purchase these cards and wait for the next algorithm to be released. But even if the next algorithm turns out to be only as profitable as GPU mining, you will still have some space and power savings. Things will only go south if the OP cannot deliver his claims.

Look, ASICs will remain dominant, then followed by GPUS; FPGAs and CPUs will have to fight for third place. You can only displace GPUs if hundreds of $millions worth of these accelerators are sold to miners and it’s not going to happen anytime soon.

[DigitalCruncher]

Any comments on the quality/portability of the athena code ?

Athena code seem to be good. Although it's interface is very narrow so you will lose a hashrate by using it.

[senseless]

Looks like Bittware reduced their standard warranty for the "crypto" models.  I believe they typically have 1 yr mfg warranty on their products.  Kind of disappointing the warranty was reduced that much for a $6k product.

Yes, a 90-day warranty raises a LOT of concerns.  Has Bittware done stress testing and has some reason to believe the failure rate ramps up after that timeframe?  I could understand a 90-day warranty on support issues with a 1-year hardware warranty.  

I was in the process of arranging my finances to purchase a few cards, but if the manufacturer has that little confidence in their product, they're sending me a message I can't ignore.  Even the Chinese manufacturers (with their notorious quality) give 180-day warranties.

Just for some background, we have been shipping this board for 1.5 years now, and have had almost no field failures.  It has gone through full validation, including thermals.  We manufacture in the US using a Tier 1 contract manufacturer. Our standard warranty is 1 year, we have some products with 3 year warranties, and some customers who pay for extended warranties.  We also generally provide unlimited support, as our normal customer is one who buys 10s or 100s (or even 1000s) of boards.  We have been building high reliability boards for decades, with one of our boards being used as part of the traction control system on the TGV (and all Alstom train and subway cars) for well over 10 years.

When this mining opportunity came up, we looked for ways to provide a deep discount as we know this community is very cost sensitive.  We also have concerns with the way that these boards will be used for mining, mostly from an ESD point of view.  Our normal customers install these boards in servers, and are in properly controlled environments with ESD protection taken very seriously.  So we lowered the warranty to reduce cost, and for fear of the boards being abused.  Perhaps we went too far lowering it to 30 days.  We also limited support, however, our normal support includes helping people develop their FPGA code on our board.  In this case you are using code already developed.

Note that we do have onboard monitoring for temperature, voltage, and current, with thresholds set to shut down the board if it appears to be entering a situation that would cause damage.  We provide a utility (console and GUI versions) that you can use to talk to this onboard controller via the USB and see what is going on with all the sensors.  Unlike GPUs and CPUs, FPGAs have no inherent thermal throttling, unless the FPGA developer builds that into their logic, which is very rarely done.  FPGAs also burn more power as they get hotter, so from a power efficiency point of view, it is better to keep them cool.

As for cost we presented, while I know it seems steep for this community, this is deeply discounted from our standard price, taking into account the reduced warranty and support.

From what I can tell, Xilinx has a 90 day warranty on their devkits, which is what the VCU1525 is.  I'm not sure how long they will sustain that low cost, we know of smaller board manufacturers that pay more than that for just the VU9P chip itself.

What is the VCCINT max amp supply on the XUPP3R? Seems like you only have enough power capacity for a maximum of around 130A vccint draw?
XUPP3R, 6Pin + PCIE = 150W
VCU1525, 8PIN + PCI-E = 225W

Ya, unfortunately you guys are bilkers aka 'how much you got?'. Though, not to be unexpected when your primary supplier and working relationship is with Xilinx. We'll call that 'trickle on' economics. I contacted you and recieved pricing almost a year ago. Was ready to jump on 50+ boards and you wanted $10K+ a piece   . I've negotiated rates lower than the cost of the VCU1525 and the MOQs aren't that high to go sub 4K. Please don't be disingenuous in an attempt to increase your bottom line. The only people who are paying more than the VCU1525 for chips are those with tiny MOQs (1-100). The lowest anyone paid for these chips is Amazon and Tencent. If I know what they paid, I'm sure you know what they paid. Which means, We both know the retail price of these boards should be around $3,000 with $4000 on the high end.

I was happy as a bouncing bean the day Intel bought Altera. They will finally open up the FPGA market to it's full potential. Do you have any idea how much better the ENTIRE PLANET would be if you stopped these manipulative tactics and started to push the product? Do you understand that data processing servers could be taken offline down to a fraction of their current number by using FPGA as a coprocessing card for the applications of that server? That the carbon footprint of ALL DATACENTERS COMBINED could be dropped to a mere fraction of what is is currently? I've got about 20KW of hardware sitting in racks right now processing apache, mysql, php, etc, things that could be processed in 1/10th the time at 1/10th the power on a FPGA. Who wouldn't want to get rid of 20 web servers operating at 300 watts each in favor of 2 300W servers with 1 100W fpga card each?! Intel can see what clearly Xilinx has been unable to see. And, we won't have to deal with a bunch of anti-competitive, anti-trust corporate thugs. But hey, why sell 10,000,000 pieces at $1,000 each ($50 prod cost) to improve the planet when you could manipulate the market and sell 1,000,000 for $10,000 each ($50 prod cost) to improve your annual bonus?

It burns me up seeing one of the most powerful computing devices ever conceived of being used as a glorified network card.

You'd be really surprised at how many products you'd move if you offered a few firmwares to accelerate sql / web / java application processing. Maybe then you can get your MOQ up to what amazon's was/is and become the supplier of a $3,000 co-processing card. Example: https://swarm64.com/how-it-works/

[Dr_Thrax]

Is there any information about this rig if it has been tested in another algorithm like equihash, dagger, lyra2rev2, nist5, skein? Thanks.

[jasemoney]

If you are data center guy such as myself, then you may prefer these server cards.

Lead time is 2 weeks:
https://www.xilinx.com/products/boards-and-kits/vcu1525-p.html#hardware

I do not trust PCIe risers with $4K hardware, there are some brave souls out here.

I got the same lead time in EU. Can you recommend any rack server for these? I guess they won't fit in normal 2U servers like HPE Proliant as most spec only one full width full height card as the board is double height if I understand it correctly.

I emailed the guy linked thru OP's post for VCU1525 through Avnet He said its the same price for active or passive and lead time is 4-8 weeks.

For a workable case, not sure if they ship to EU but I have two of these 6 card cases for gpus at a local my datacenter: https://www.miningrigs.net/product/gray-matter-v3-1-gpu-case/ they have great airflow. They take up 4U of space, and some rails to sit on are sold seprately. I have a mismash of different cards and i notice the blower style cards run hotter in this than the open fan gpus, so id probably go for passive fpga and buy another of these cases for them. I would probably try to modify the airflow so it was forced to go through the cards rather than through & around, though consideration for the back of board cooling would be in order.

[yrk1957]

I was browsing the Baikal X10 thread and what the OP said there, this was March 12:

I am building my own FPGA rig and after a detailed analysis of all current algorithms, I can say that all algorithms will soon be mined by ASIC/FPGA except equihash and ethash.  Equihash will fall to ASIC's in about 6 years when we can get 144MB of SRAM on a single chip.  Ethash will not fall to ASICs ever, in my opinion.  The 2GB scratchpad is just too big.

Cryptonight and Cryptonight-Lite are mildly ASIC resistant.  This means that even a dedicated unit like the Giant-N cannot 'crush' GPU's.  It can beat them marginally in terms of ROI.  But for other coins, ASICs crush the GPU's by hundreds of times.

Eric

Two things to to note: Equihash falls to ASIC in 6yrs and Cryptonight ASICs will beat GPUs only margnially.


So a couple of days later, Bitmain announced X3 with 10x Cryptonight performance than Giant-N, and equivalent to 100 odd Vega-64s.

And 2 months later, we have the Z9 Equihash ASIC.


And now I have my doubts on OPs assessments...


Edit:

This is proposed performance for Phi on VC1525:

Phi1612 (Luxcoin, Folm): 5.2GH/s (650MH/s per card x eight) ($456/day at Apr-30 prices)

So 650MH/s = 20x 1080Ti

That should be well within an ASIC range, if and when one emerges.


Edit2:

While the OP maybe an FPGA expert, he does not appear to have much experience in what ASICs can do. Bitmain has!

[HardwareCollector]

This is proposed performance for Phi on VC1525:

Phi1612 (Luxcoin, Folm): 5.2GH/s (650MH/s per card x eight) ($456/day at Apr-30 prices)

It was a claim not an estimate.
You are allowed to make mistakes with an estimate, but you have to backup your claims.

People asked me to make a new thread on this topic, so here it is.

Here are some pics and video of my 8 x Xilinx VCU1525 rig.  Each VCU1525 card has one Xilinx VU9P Virtex Ultrascale+ FPGA.  Hash rate for the whole rig combined is:

Keccak (Smartcash, Maxcoin): 136GH/s (17GH/s per card x eight) ($160/day at Apr-30 prices)
Tribus (Denarius, Virtus): 16.8GH/s (2.1GH/s per card x eight) ($304/day at Apr-30 prices)
Phi1612 (Luxcoin, Folm): 5.2GH/s (650MH/s per card x eight) ($456/day at Apr-30 prices)
Skunhash (Various coins): 10.4GH/s (1.3GH/s per card x eight) ($261/day at Apr-30 prices)

Those yield around US$20-$57 per card per day ($160-$456 per day for the rig).  Each VCU1525 card costs $4000, or $32K for the whole rig.  At $160-$456 per day, ROI is 70-200 days depending on the algorithm.  I'm not the only one mining with these cards.  Apparently some guy in Germany is getting 64KH/s with Cryptonight-V7 on the same VCU1525's, earning him over $100 per day per card or $800+ per day for a whole rig.

[yrk1957]

It was a claim not an estimate.
You are allowed to make mistakes with an estimate, but you have to backup your claims.

That's not my point. (As I said he "proposed" this performance.)

The assessments of OP around Crypto/Equihash are not in the mistakes category, that seems to be a lack of knowledge on ASIC side, IMHO.

Note that I have Zero and Zero experience on FPGA and ASIC.

[abudfv2008]

Is there any real proof of declared speed?
Just point it to yiimp pool and all could see real hashrate. 

[HardwareCollector]

The assessments of OP around Crypto/Equihash are not in the mistakes category, that seems to be a lack of knowledge on ASIC side, IMHO.

FPGAs and ASICs are very closely ralated, and FPGAs are generally used to prototype ASICs. But the statements below made by the OP shows that he has been removed from the industry for far too long.

I really wish people would stop calling FPGA's ASICs.  Baikal (so far) has only made FPGA mining rigs.  There is a huge difference between an FPGA and an ASIC.  An FPGA is not that much different from a GPU.  Anyone can get one.  Anyone can buy a Zynq FPGA board from digikey for $89-$199, or a higher end one for more money, and if you take a little while to learn how to program it, you can hash any algorithm except equihash & ethash.  Furthermore, your ROI will be better than a GPU in almost every case, in some cases dramatically better (as Baikal showed with the X10 and Giant-B).

If you want evidence of the OP's claims, there's none for now and you will have to wait until May 30th.

[abenhamo]

Hello,
Can R0land show us an evidence of the reached hashrate on CNv7?
Thank you

[Exchange Email]

very interested in this project, i sent an PM

[yslyung]

just my 2 cents...

seeing is believing

[cryptaioracle]

I learned about this post and read in detail. It has become very inspiring, as I am both a crypto mining, AI and energy enthusiast (and professional). There are looming issues around the sustainability of PoW crypto economies, concentration of power in crypto economies, regulation of crypto economies and even availability of power for not only crypto security but also for other high density advancements in technology driven by AI.

For a long time I have felt that there needed to be a more distributed solution to secure the network.  Many limitations to GPU mining including power consumption and regulation (even as to individuals but also commercial operations) and simply a huge difficulty in building and maintaining mining rigs that are preventing a broader adoption. Normal people simply can't handle it.  I have tons of friends who are like "I would love to have one of those in my living room, I don't care how much it makes but it looks so cool!" I meet software developers that come into my office and see my mining rigs and don't even know what they are. We think everyone in the world knows everything about crypto.  But they don't.

This FPGA solution can answer many of those problems, by reducing the amount of power, space and cooling required to maintain blockchain security in PoW systems.  Prior to the newer and larger FPGA products this hasn't really been possible.

This has inspired me to consider developing a consumer grade, multi-purpose work station with specific capabilities needed for FPGA mining and deep learning networks. A "Personal Accelerator"

Link to sample marketing document I created for the project: https://www.dropbox.com/s/7w6ibv6q6o5x21k/algorithmic.energy.pdf?dl=0

As many people have commented, an open source, community based development approach is probably best.  The question remains how to compensate early developers to get it going. Because the FPGA firmware will be hardware specific, this should be easy for hardware makers to do based on cost recovery from hardware sales. I believe that the developers should be compensated for time and costs as any engineer is, but that the growth and maintenance of the available firmware libraries can be built on open source networks and a dedicated (even decentralized) organization. 

In a consumer grade product with multiple potential utility streams, it would be possible to leverage traditional financing solutions and with sufficient revenue from mining or cloud computing operations would be net revenue positive from the first day of ownership.  By exponentially expanding the potential network, outside of government regulation and concentration of power, it would be possible to overcome both the GPU and ASIC hashing capacity. 

What better place to open this can of worms that here.  Looking forward to thoughts and comments.

Initial questions would be:
1. Optimal selection of hardware based on cost/performance/packaging.  Currently looking at Alpha Data 8k5 with Xilinx KU115 core or comparable packages. Open to any suggestions, however the VU9 does seem a bit much for this kind of solution, the FPGA package needs to come in the $3K-$4K range tops and I don't see Xilinx doing it, even at volume.  A data center will pay much more than that, why would they lower their pants?  They won't... 
2. Support from OP and/or other thread members in working on the open source, based on below T&C assumptions?
3. Should there be a "maintenance model" in which the Open Source Org captures a portion of the revenue to maintain and grow support for new algorithms and uses (including deep learning networks), OR should there be a fixed fee charged upfront (included in hardware purchase price) for a "support term" and then an annual support fee when that term expires? I personally prefer the latter, much simpler.
4. Any ideas on how big of a support team would be required to maintain the org?  Would 3 full time developers x $150K/year salary+ 1 full time QC x $100K/year salary overhead $50K/head = $750K/year be sufficient?   
5. Would people buy this hardware for the suggested price range $5500-$6500? Would they finance it for 24 months for $265/month (including support and warranty)?  Would your friends or family do the same?
6. Anybody interested in investing or partnering in the hardware company? I am currently already looking at MFG solutions and supply chain and have most of that scoped out for volume production, the main part missing would be the firmware and algorithm development and QC. I have been building hardware, sensors, and software (including SaaS) for the last 18 years of my life.  I think we can handle this and put a good team together. 
7. What do people think of a "trade in" option, where you can get a credit value for exchanging your current GPU mining rig? We could reclaim existing GPU's, set them back to factory specs and then release them back to the gamers for reasonable prices that they so desperately have needed.

Although I appreciate answers from everyone, if someone is interested in a closer relationship I would only be interested in working with serious, professional people. This is not intended to be a hobbyist solution, but a real commercial, consumer grade product run by competent, experienced and professional management.  If you have less than 5-10 years of experience in the field and don't have credible professional references as an entrepreneaur, investor or engineer, then this is probably not an opportunity for you at this time (not to be unfair, just realistic). 

[cheebeez]

FPGA data + 500k$ = ASIC
Good luck 4% guy

[yrk1957]

FPGAs and ASICs are very closely ralated, and FPGAs are generally used to prototype ASICs.

FPGA data + 500k$ = ASIC
Good luck 4% guy

Wait Wait.

You guys are saying that with the FPGA and bitstream code OP has, someone with deep pockets can actually produce an ASIC from it?

[jasemoney]

FPGAs and ASICs are very closely ralated, and FPGAs are generally used to prototype ASICs.

FPGA data + 500k$ = ASIC
Good luck 4% guy

Wait Wait.

You guys are saying that with the FPGA and bitstream code OP has, someone with deep pockets can actually produce an ASIC from it?

thats how it works... well youd need to know how the fpga layout was not just have the bitstream i assume.
take the logic you can program into an fpga and hardcode it onto a chip and throw away everything else there is your asic chip. of course the $500k cost mentioned is referencing the cost of tapeout (unclear actual cost) to actually create the chip you have logic for.