Topic: DIY FPGA Mining rig for any algorithm with fast ROI - page 83. (Read 99472 times)

We are low level RTL here, so I haven’t messed with much of the HLS side. The big issue with FPGAs is people used to buying graphics cards that are all commoditized close to a low MSRP don’t understand that the FPGA market is nothing like this. The traditional market for these things was Avionics, Defense, Government, Nation States, etc. the price on digikey is no where near the actual price for someone building their own hardware, and the dev kits are expensive because they come with every bell and whistle you might want to use with a particular chip and they are low low volume. This is changing a bit but these VCU1525 dev accelerators are no different than Nvidia’s P100/V100 PCIe card’s or AMD’s old FirePro, now Radeon Pro multi-thousand dollar cards. They don’t cost 10x more to make, the target market is just far less cost sensitive.

The gap between ASIC and FPGA is also much narrower now than it was a few years ago.

If a market emerges to consume 100,000 FPGA accelerators for mining, someone building their own hardware with a good relationship with Xilinx or Intel can and will drive the cost down at least an order of magnitude.

@GPUHoarder

What has been your experience with the latest HLS tools from Xilinx? A few years ago, the performance of their sample Bitcoin miner was atrocious. A $2K board with the throughput of 80 MH/s, not sure why they even included that sample.

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.

I was thinking PCIe spec backplanes, but I see what you’re referring to. I would be surprised if the edge connectors electromechanical could support higher than 16 per transceiver, they’re a pretty loose fit. I’ve seen how much engineering goes into things like sea-ray connectors but as you said BER may remain acceptable for losses crypto mining. I also forgot PCIe is bi-directional as well, so 16 TX + 16 RX pairs as much as ~400Gbps each way in a ring. still 1/16th of what would be available with a dedicated 128 transceiver design, and those parts are in the ballpark of a VCU1525 in quantity. Take a look at some of the ASIC verification setups based on 4+ VUP parts. HTG I believe has some.

OK, so both Xilinx and Bittware boards have 16 xcvrs hooked to PCIe edge connector.

I mentioned "PCIe passive backplane" to emphasize that to use that we wouldn't need to obey the full protocol stack, including BIOS discovery & negotiation, transport layer packetization, etc. If the passive backplane contains only cards running "our" project then we could violate many if not most constraints in the official protocol.

All that really matters are the mechanical fit and signal margins on the electrical connectors. Given the simplicity and low cost it wouldn't be a problem to even cut some traces on the backplane PCB to make it electrically independent and improve noise margins by cutting out unnecessary copper.

Admittedly, I haven't played this game with PCIe devices, but in the past with older technologies all the way to PCI and PCI-X. Additionally, mining is essentially a lottery, so it can profitably tolerate bit error rates that wouldn't be acceptable in conventional applications. I get a sense that most of the devices utilizing QSFP28 are designed for either longer distances or higher reliability than what would be required here.

Or maybe you just thought about PCIe 4.0 motherboards instead of passive backplanes?

@GPUHoarder
[/quote]

I wouldn’t say an all out attack on GPUs by FPGAs. A good portion of my work has been on FPGA “accelerators” that augment GPUs. GPUs still have way more memory bandwidth than FPGAs after the cache-sized blocks. I have developed some interesting FPGA coprocessor that replace inefficient GPU calculations with FPGA work but still do the memory hard loops on GPU, for overall hashrate and power gains. I expect that hardware to be ready midsummer. Remains to be seen if it will still be viable.
[/quote]

Can you share some more info/specifics on this? I find this interesting and something that has crossed my mind a few times

hey I am a geek,but I have a degree in accounting along with a lot of economics.

So this brings out a very interesting economic model.

My core belief is 50 to 100 million miners minimum or adoption struggles.
Gpu's were the easiest way to have 50 to 100 million miners.

I built gaming rigs since 2004 every gamer I know mines a bit when they can not game. Cant see killing that off being good for crypto coins.   Just like I can't see  POS working for coins.

I can see varitions of the op's idea including his actual idea being viable as a side base.

NVIDIA makes the guts for Switch and AMD does Xbox and PS so gamers moving to consoles maybe isn't that bad. Gamers moving to mobile platforms might be more of a problem although I'm not sure if anyone would really say "dang this 1080 is too expensive so I'm gonna play Candy Crush instead"


However they're exposed to those crashes and booms already. So I think the point Phil was making is that NVIDIA/AMD would actually have more control if they developed their own algo or otherwise embraced crypto in a more "official" way.

I wouldn’t say an all out attack on GPUs by FPGAs. A good portion of my work has been on FPGA “accelerators” that augment GPUs. GPUs still have way more memory bandwidth than FPGAs after the cache-sized blocks. I have developed some interesting FPGA coprocessor that replace inefficient GPU calculations with FPGA work but still do the memory hard loops on GPU, for overall hashrate and power gains. I expect that hardware to be ready midsummer. Remains to be seen if it will still be viable.

Right, the GPU dump on crypto crash is one side of the bad scenarios for Nvidia (the other side being extremely high crypto demand which drives gamers to consoles and they loose a large chunk of their GPU gaming business).

But Nvidia does not control whether crypto crashes or booms. They have to have a strategy keep their base (GPU gaming/Datacenter) market intact in either of these scenarios. I don't think having their own coin is a viable strategy for these scenarios, it will just drag them deeper into the crypto world.

I am here also interested in these FPGAs mostly as a hobby. And I suppose many others are in same way. Whether large farms adopt them, will probably happen only when these FPGA's are proven. So nothing much may change in next 3-6 months.

This is why I bookmarked thread.  I want to see how this develops.

I can see some people have technical interest in it.

I can also see why op wants 4% fee not 2% or 1.5%  as  he fears  this won't last for him.

I respect that fear it is a healthy one.  Looks like this year will be an all out attack on gpus  via asics and fpgas  should be fun to watch it happen.

I won’t say it’s impossible, but I would be really genuinely surprised. 32MB / hash of total bandwidth (read + write) is needed, and 2MB or so of stashes per hashcore.

You have 1280 URAM blocks of 288kb by 72 bit interface dual ported in the biggest configuration .That’s an incredible amount of internal bandwidth but you can only store 23 or so simultaneous Cryptonight7 2MB blocks in that. The absolute biggest part (which isn’t on the 1525 board) has 360Mbit URAM, 96Mbit BRAM, and 48Mbit Distributed RAM, holding a theoretical 63 MB of pipelines, assuming you didn’t need a single bit of that for the rest of your logic (you do).

The external memory at say 4x64 DIMMs @2666 is only 85GB/s, or 2.6 KH worth of bandwidth with a perfect access pattern.

Even if you could imaginarily use all 2000+ balls on the FPGA for 2666 MT/s DDR style  speeds you’d still only clear 20KH against external memory and that isn’t even real bandwidth.

 Even if you took the biggest part with 128x32 Gbps transceivers to SERDES memory you’d only have 16kH limit from bandwidth.

Unless you break the algorithm itself, there’s no where to find the bandwidth + storage space for 64khs on a single FPGA.

not for mining use, its 30 days (or 90?) they reduced it from their standard warranty period. i believe they said you can buy extended warranties however. thier main concern was DIYers ignoring ESM precautions and abuse due to ignoring proper installation (heat, power quality, improper installation etc).

still these are industrial units with a proven reliability record, in wide use in industry (so bittmex says, and i have no reason to doubt them atm) , not bitmain trash. while i am a bit disappointing, i still would buy.

their monitoring software monitors a lot of sensors to keep it in the safe zone, which is the main reliability point imo.

PCIe, even 4.0 for which backplanes are not yet available, is 8 or 16Gbps per transceiver vs 28 Gbps in the QSFP28s

I wasnt referring to 1525s though, I was speaking about other boards with 56+ transceivers exposed.

If you need more FPGA to FPGA connectivity, the XUPP3R also has an onboard expansion connector with 20 transceivers, and a standard cable that can connect two boards together.  So 4x 100Gbps QSFPs on the front, and another 5x 100Gbps on an expansion connector.  That means you could daisy chain with 400Gbps FPGA to FPGA.

Well 22khs is pretty descent, the guy who claims VCU1525 gets 64khs is R0land

Yep  the big fear for them is a complete crash making a glut of 300-400 dollar 1080ti's    even AMD has this issue.

If I am a gamer  and this seller on ebay


https://feedback.ebay.com/ws/eBayISAPI.dll?ViewFeedback2&userid=philipma1957


is selling 350 dollar 1080ti's

{note the seller is me I am not selling any 1080tis}

I would buy that 1080ti  knowing I have an ebay return.  This happened with last the crash. Except it affected AMD a lot more then Nvidia.

This time both builders will be affected.

So like it or not both AMD and Nvidia  will have to counter move  on Asics and FPGAs.

I think the op has a good idea here  I just caution people about countermoves.

Once again  if my  Nvidia-Amd coin of the month  is invented  with 36 algos ready  to be placed in on the first of each  month for my mining card with a 3 year warranty  how fast does the fpga get reprogrammed?  Also how long is its warranty?

GPUS with 3 year warranties have  a place in a low power operation.  My solar array deal with buysolar is 'free' power  So the gpus would earn for  as long as they last.

Does an fpga have a 36 month warranty?

They're in the game whether they want it or not. NVIDIA makes the P series specifically for mining. They also acknowledge the revenue albeit as "difficult to quantify".

https://www.nasdaq.com/aspx/call-transcript.aspx?StoryId=4144915&Title=nvidia-s-nvda-ceo-jen-hsun-huang-on-q4-2018-results-earnings-call-transcript

I can wait  to watch it shake out.  I am down from 60x 1080tis to 18 of them. The 18 make a  small profit.

For now I am more interested in LTC  as it has a ton of asic hash and is mature.

I bookmarked this thread.   As I would most likely order one of these once it is mature.

I also know if asic and fpga trash gpu profits.  There will be a glut of cards on the market if this happens Nvidia and AMD  won't sell much of the new cards they make.
I have to think  they will do something about that.

i wonder that also. its one thing to put out custom mining cards, but for either company to actually design a coin, figure a use for it (store of value, utility, whatever) its algos, and get it traded on exchanges seems pretty risky as now they are directly involved in "printing money" so to speak. thats way out of their normal business.

maybe a spinoff company? dunno.

and what if the coin bombs? just because amd or nvidia makes it doesnt mean it wont be just another shitcoin. might be a lot of r&d down the tubes.

they might be able to (unofficially? but thats risky too if caught) support some devs of an existing or new coin. or be an advisor or something.

i can see devs moving to algos that need 10 gb or more (1080ti class memory) in the short term but even that wont hold back asic/fpga for long, looks like its getting much easier to add large chunks of ram to fpga, the uxppr3 has 4 rdimm slots in it. plus its ability to take (the super expensive and stupid fast) qdrii+ memory.

im still gonna risk a fpga. i must admit it will be mainly for the fun and to play with it. i dont really see it as a super weapon or major money maker though. but i dont see it as being super terrible, i would think it would eventually make its money back but years vs months? who knows.