DIY FPGA Mining rig for any algorithm with fast ROI

[heavyarms1912]

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?

I believe you need the low-level source code (behavioral logic, transistor-gate level) to make an ASIC out. 
Or you would have to reverse engineer the bit-stream which would also require complete knowledge about entire FPGA layout and schematics?

[GPUHoarder]

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?

I believe you need the low-level source code (behavioral logic, transistor-gate level) to make an ASIC out.  
Or you would have to reverse engineer the bit-stream which would also require complete knowledge about entire FPGA layout and schematics?

There isn’t really much magic to the RTL for these things. Most of the performance gains are more about working around the nuances of the FPGA design and finding hardware where someone has laid out all the needed components (memory bandwidth and/or low latency memory, multiple chips with high speed interconnect, etc. I do wonder how the the OP intends to secure the bitstream if he distributed it. It would be pretty easy to write your own miner, and you can only encrypt the bitstream if it is your own hardware and you burn eFuses or similar.

Also with FPGAs and dev boards you are usually given very detailed manuals and all the schematics.

Anyone equipped to do ASICs could implement the algorithms themselves, it’s all about the capital - and accepting that all that cost is wasted if the algorithm is changed. For a lot of modern algorithms it is more about assembling the memory components than even the ASIC cores.

The reason some of us are long on FPGAs is they have the potential to be lower cost, much lower power, and higher performance than GPUs for algorithms that choose to resist ASIC development (aka change regularly). There is very little a GPU can do to gain advantage on modern (upcoming) FPGAs in that world. The transition to HBM as standard and Intel ending the artificial high price racket are the only barriers. It doesn’t take that long to stand up working RTL for an algorithm change. Personally I’m thrilled if Crypto pressure ushers in a whole new era of high performance computing.

[senseless]

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?

They could make a structured asic which doesn't have the same level of performance / energy efficiency gains as a 'real' asic (cell / custom, etc). It would lower cost a bit depending on quantity and maybe 15-20% performance/efficiency improvement.

[nsummy]

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/

I don't know if I would exactly depend on Intel as the savior.  There are probably thousands of products listed on their ARK page that are impossible for a consumer to get.  Anything that they come out with that has great performance and a reasonable oem price seems to be exclusively sold by Supermicro at wildly inflated prices.

[Amph]

just think about it, it's not that crazy as may sound like at first glance

one of this is just 30 gpu 1070, at the cost of 10 o f them(you can find a 1070 for 330 or lower now so it's up to 12 of them), so it's 3:1 at best, but the new nvidia gpu will be 100% faster than a single 1070

this mean that this vs the 1170 will be only 1.5:1, 50% faster or lower...nothing special

https://www.techradar.com/news/nvidia-geforce-gtx-1180-leak-shows-that-its-faster-than-a-titan-xp

[Ginzink]

just think about it, it's not that crazy as may sound like at first glance

one of this is just 30 gpu 1070, at the cost of 10 o f them(you can find a 1070 for 330 or lower now so it's up to 12 of them), so it's 3:1 at best, but the new nvidia gpu will be 100% faster than a single 1070

this mean that this vs the 1170 will be only 1.5:1, 50% faster or lower...nothing special

https://www.techradar.com/news/nvidia-geforce-gtx-1180-leak-shows-that-its-faster-than-a-titan-xp

That Nvidia will release a GPU that is 100% faster than 1070 for 330 USD (which is you price in the comparison) is quite unrealistic
This generation we had quite a boost, yet the 1060 is only about 12% faster than a 970 as an example.
And dont forget the difference in power consumption! I would love a single FPGA instead of a 1080 ti rig, but have to see that it works first. Lead time in Norway is 12 weeks though

[Amph]

just think about it, it's not that crazy as may sound like at first glance

one of this is just 30 gpu 1070, at the cost of 10 o f them(you can find a 1070 for 330 or lower now so it's up to 12 of them), so it's 3:1 at best, but the new nvidia gpu will be 100% faster than a single 1070

this mean that this vs the 1170 will be only 1.5:1, 50% faster or lower...nothing special

https://www.techradar.com/news/nvidia-geforce-gtx-1180-leak-shows-that-its-faster-than-a-titan-xp

That Nvidia will release a GPU that is 100% faster than 1070 for 330 USD (which is you price in the comparison) is quite unrealistic
This generation we had quite a boost, yet the 1060 is only about 12% faster than a 970 as an example.
And dont forget the difference in power consumption! I would love a single FPGA instead of a 1080 ti rig, but have to see that it works first. Lead time in Norway is 12 weeks though

will be 500 at launch( so 15 1170 vs this board will be $4000 vs $7500)as usual and 700 for the 1180 which is arguably as a fast as this fpga board for the same cost and same consumption with some tweaking

a 1170 will consume 100 watt, tdp of a 1180 which is a strongest card is the same as 1070ti which consume only 100, so a 1170 will consume less probably at lower tdp

[m5]

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?

Maybe. But that ASIC would be static and could not be modified to other algos. FPGA can be pre-programmed quite easily if coins do changes to their algos. You won't have a doorstopper/heater from FPGA overnight as you have with cryptonight ASICs now.

[GPUHoarder]

just think about it, it's not that crazy as may sound like at first glance

one of this is just 30 gpu 1070, at the cost of 10 o f them(you can find a 1070 for 330 or lower now so it's up to 12 of them), so it's 3:1 at best, but the new nvidia gpu will be 100% faster than a single 1070

this mean that this vs the 1170 will be only 1.5:1, 50% faster or lower...nothing special

https://www.techradar.com/news/nvidia-geforce-gtx-1180-leak-shows-that-its-faster-than-a-titan-xp

That Nvidia will release a GPU that is 100% faster than 1070 for 330 USD (which is you price in the comparison) is quite unrealistic
This generation we had quite a boost, yet the 1060 is only about 12% faster than a 970 as an example.
And dont forget the difference in power consumption! I would love a single FPGA instead of a 1080 ti rig, but have to see that it works first. Lead time in Norway is 12 weeks though

will be 500 at launch( so 15 1170 vs this board will be $4000 vs $7500)as usual and 700 for the 1180 which is arguably as a fast as this fpga board for the same cost and same consumption with some tweaking

a 1170 will consume 100 watt, tdp of a 1180 which is a strongest card is the same as 1070ti which consume only 100, so a 1170 will consume less probably at lower tdp

Did I miss the detailed performances spec, price, and power consumption figures on the 1170 getting published?

[zer0k]

just think about it, it's not that crazy as may sound like at first glance

one of this is just 30 gpu 1070, at the cost of 10 o f them(you can find a 1070 for 330 or lower now so it's up to 12 of them), so it's 3:1 at best, but the new nvidia gpu will be 100% faster than a single 1070

this mean that this vs the 1170 will be only 1.5:1, 50% faster or lower...nothing special

https://www.techradar.com/news/nvidia-geforce-gtx-1180-leak-shows-that-its-faster-than-a-titan-xp

That Nvidia will release a GPU that is 100% faster than 1070 for 330 USD (which is you price in the comparison) is quite unrealistic
This generation we had quite a boost, yet the 1060 is only about 12% faster than a 970 as an example.
And dont forget the difference in power consumption! I would love a single FPGA instead of a 1080 ti rig, but have to see that it works first. Lead time in Norway is 12 weeks though

will be 500 at launch( so 15 1170 vs this board will be $4000 vs $7500)as usual and 700 for the 1180 which is arguably as a fast as this fpga board for the same cost and same consumption with some tweaking

a 1170 will consume 100 watt, tdp of a 1180 which is a strongest card is the same as 1070ti which consume only 100, so a 1170 will consume less probably at lower tdp

You are dreaming if you think the new 1180 will be as fast as this card.
The 1180 will be an incremental improvement over the outgoing 1080, and even if it was faster than a Titan it will still be far slower.

Depending on the algorithm, we're talking about an FPGA that looks like it can be equivalent to 10+ Titan's!

[Techman34]

Hello,

very intresting project. Sorry for the stupid question but on your video - the pci riser are not connected to the motherboard ? How is the pci fpga card bus communicating with the motherboard ?

Thanks

Normal USB ports

Sorry if it's already been answered.  Is the USB controller a bottleneck for a multi-FPGA build?  On a normal motherboard you can use a USB hub to expand the USB slots but there is a bandwidth constraint on the USB controller itself.  Are a few 4 port expanded controllers needed to expand that to have each FPGA card on it's own dedicated USB controller for max throughput?  Something like the Sonnet Allegro Pro?  Thanks,

Techman34

[Chugumoto]

Hello,

very intresting project. Sorry for the stupid question but on your video - the pci riser are not connected to the motherboard ? How is the pci fpga card bus communicating with the motherboard ?

Thanks

Normal USB ports

Sorry if it's already been answered.  Is the USB controller a bottleneck for a multi-FPGA build?  On a normal motherboard you can use a USB hub to expand the USB slots but there is a bandwidth constraint on the USB controller itself.  Are a few 4 port expanded controllers needed to expand that to have each FPGA card on it's own dedicated USB controller for max throughput?  Something like the Sonnet Allegro Pro?  Thanks,

Techman34

if my memory does not change me, wrote that the speed of usb2.0 is even enough

[ilia_2s]

U R right, it require a little amount of bandwidth only to send a small hashing task and return a result.

[Kainzo]

I contacted Bittware and they are delaying the release of the crypto version of the XUPP3R.
Looks like I may have missed out.

[crypto4pizza]

I contacted Bittware and they are delaying the release of the crypto version of the XUPP3R.
Looks like I may have missed out.

Did they provide any details on current lead times?  It was 4-5 weeks previously.

[BittWareFPGATech]

I contacted Bittware and they are delaying the release of the crypto version of the XUPP3R.
Looks like I may have missed out.

Sorry, I've been meaning to make a post related to this.  We have been working with the OP to determine the optimal configuration and board for this application, as we have a few to choose from.  The XUPP3R may not be the optimal one, we have other boards, including one with larger core power supplies and even the VU13P instead of the VU9P.  The OP is working to determine the combination of FPGA size, clocks, logic, memory and power use that maximizes ROI.  We do not want to sell you boards that may not be the best fit, that is in no one's best interest.

So you have not missed out, you perhaps have saved yourself from buying a non-optimal board (we actually have not yet taken any orders or shipped any boards to miners so no one is in that camp).  You can still contact us and we will get back to you as soon as we have enough details to let you place an order.

[netmonk]

I contacted Bittware and they are delaying the release of the crypto version of the XUPP3R.
Looks like I may have missed out.

Sorry, I've been meaning to make a post related to this.  We have been working with the OP to determine the optimal configuration and board for this application, as we have a few to choose from.  The XUPP3R may not be the optimal one, we have other boards, including one with larger core power supplies and even the VU13P instead of the VU9P.  The OP is working to determine the combination of FPGA size, clocks, logic, memory and power use that maximizes ROI.  We do not want to sell you boards that may not be the best fit, that is in no one's best interest.

So you have not missed out, you perhaps have saved yourself from buying a non-optimal board (we actually have not yet taken any orders or shipped any boards to miners so no one is in that camp).  You can still contact us and we will get back to you as soon as we have enough details to let you place an order.

Interesting, so on what hardware OP was first developping ? I mean, in the first post, there is two devices listed. I hope OP didn't push your board without some effective experiments.
This delay and those explanations are quite suspicious. Either OP has working bitsream with real number on the VU9p, either he has nothing today.

My opinion is that you are quite surprised by the community huge demand, and you don't know how to face it, also in regards with your regular customers who are going to push price down once you sell similar card cheaper to miner. lack of preparation and fear of unknown

[tromp]

Sorry, I've been meaning to make a post related to this.  We have been working with the OP to determine the optimal configuration and board for this application, as we have a few to choose from.

If you could include a table showing how much (LP)DRAM and SRAM (or low-latency alternative) can be hooked up to each FPGA, as well as the resulting cost, that would be very helpful. Thanks in advance!

[Etherion]

Talk about cutting edge.BittWareFPGATech/ Christian, it looks like you and the OP is standing on the edge of the rabbet hole. I hope BiiWare can beat Baikal and Bitmain in provideing a good custom mining solution that will provide good decentralizion and diverse user base.

[senseless]

Sorry, I've been meaning to make a post related to this.  We have been working with the OP to determine the optimal configuration and board for this application, as we have a few to choose from.

If you could include a table showing how much (LP)DRAM and SRAM (or low-latency alternative) can be hooked up to each FPGA, as well as the resulting cost, that would be very helpful. Thanks in advance!

The best thing to hook up to the FPGA would be a hybrid memory cube +$500. This would get you the same level of memory performance as HBM memory. The other nice thing, the HMC has a silicon memory controller on it along with some basic logic functions that can speed up certain applications (xor, and, or).