DIY FPGA Mining rig for any algorithm with fast ROI

[s1gs3gv]

I'm pretty sure intel's plan is to flood the market. Why else would they be developing hybrid cpu/fpgas (like APUs) or why would they bother to create the CCIX interconnect? They're definitely going to flood the market.

Not exactly flooding the market with cheap ultrascale+ class products ...

https://www.anandtech.com/show/12773/intel-shows-xeon-scalable-gold-6138p-with-integrated-fpga-shipping-to-vendors

sampling now to special customers. spendy and only an arria 10.

it'll be quite a while before we're seeing affordable integrated stratix10 / x86 products that mom and pop can afford.

[1714812388]

1714812388

[senseless]

I'm pretty sure intel's plan is to flood the market. Why else would they be developing hybrid cpu/fpgas (like APUs) or why would they bother to create the CCIX interconnect? They're definitely going to flood the market.

Not exactly flooding the market with cheap ultrascale+ class products ...

https://www.anandtech.com/show/12773/intel-shows-xeon-scalable-gold-6138p-with-integrated-fpga-shipping-to-vendors

sampling now to special customers. spendy and only an arria 10.

it'll be quite a while before we're seeing affordable integrated stratix10 / x86 products that mom and pop can afford.

I think our definitions of flood the market differ. You're thinking about it from a consumer level -- It will eventually be that way -- But not in the next 12 months. If intel had the same practices as Xilinx, the Xeon 4116 system I just built would have cost me $100,000... Not $6,000... And I would have had to design my own motherboard, because they wouldn't have allowed me to produce their reference design... I don't think it's unreasonable to expect a $500-1000 unit cost on a large Stratix or Ultrascale+ part with HBM2 once production is moving 100K units a month. If Xilinx has their way, that part would cost $5,000-10,000 at "volume". If Intel increases their market share by expanding the market, and making these chips available at the enterprise level for hardware acceleration of databases, web servers, etc.... It's not unreasonable to think they could move 100K units a month at a $500-$1000 cost.. Amazon is currently leasing Xilinx fpgas. Do you think AWS would be interested in including FPGA logic in every single one of their servers? What about tencent? What about Baidu? What about every other provider on the planet? There's a storm on the horizon and it's going to be beautiful to watch.

The value of these devices in the server markets is in the 100s of billions. Far beyond their value in crypto markets, defense, etc, IMO. Xilinx has been very short sighted in their attempt to make profit. The world needs low cost FPGA. Not having it is preventing major advancements in computing.

[Iamtutut]

Can you give anymore details on what exactly you are going to try to get going with Intel? Are you talking Stratix10 HBM? Or some form of prototype?

Yes, the Stratix 10 parts are equivalent to the Ultrascale+ parts. I believe we'll be able to get the same or better performance for a much lower price point. If FPGA are really going to be viable in the long term to displace GPUs the cost has to be reduced in a significant way. $3995 is ok, right now, while margins are high... But if we really wanted to displace GPUs completely both the cost / perf and power / perf will need to be better than GPUs. I'm hoping that we'll eventually (6-12 months) be able to sell a $1000-2000 board with a stratix 10 or ultrascale+ part that has HBM.

What about this very critical point of view about Intel Stratix 10 HBM ?
https://www.semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-dont-hold-water/

[Etherion]

Whoosh

Are you going to share or just boast?

[GPUHoarder]

Can you give anymore details on what exactly you are going to try to get going with Intel? Are you talking Stratix10 HBM? Or some form of prototype?

Yes, the Stratix 10 parts are equivalent to the Ultrascale+ parts. I believe we'll be able to get the same or better performance for a much lower price point. If FPGA are really going to be viable in the long term to displace GPUs the cost has to be reduced in a significant way. $3995 is ok, right now, while margins are high... But if we really wanted to displace GPUs completely both the cost / perf and power / perf will need to be better than GPUs. I'm hoping that we'll eventually (6-12 months) be able to sell a $1000-2000 board with a stratix 10 or ultrascale+ part that has HBM.

What about this very critical point of view about Intel Stratix 10 HBM ?
https://www.semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-dont-hold-water/

Both Xilinx and Intel are failing to sample or ship HBM chips announced more than six months ago. I don’t think that HBM is coming at mainstream prices in FPGAs anytime soon - even AMD and Nvidia are moving away from it in consumer products.

[GPUHoarder]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

[senseless]

Can you give anymore details on what exactly you are going to try to get going with Intel? Are you talking Stratix10 HBM? Or some form of prototype?

Yes, the Stratix 10 parts are equivalent to the Ultrascale+ parts. I believe we'll be able to get the same or better performance for a much lower price point. If FPGA are really going to be viable in the long term to displace GPUs the cost has to be reduced in a significant way. $3995 is ok, right now, while margins are high... But if we really wanted to displace GPUs completely both the cost / perf and power / perf will need to be better than GPUs. I'm hoping that we'll eventually (6-12 months) be able to sell a $1000-2000 board with a stratix 10 or ultrascale+ part that has HBM.

What about this very critical point of view about Intel Stratix 10 HBM ?
https://www.semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-dont-hold-water/

Both Xilinx and Intel are failing to sample or ship HBM chips announced more than six months ago. I don’t think that HBM is coming at mainstream prices in FPGAs anytime soon - even AMD and Nvidia are moving away from it in consumer products.

As far as I'm aware they are at least in some limited fashion shipping the stratix 10 mx series. I'll confirm next week for sure. But I do remember hearing about problems they had in 2017 and how they wouldn't be resolved until 2018 because they had to redesign the stratix 10. I believe what we are now seeing is that redesigned chip. As far as I know, altera had never planned 58G serdes in the stratix 10 -- unless I missed it in the literature of the time.

[Iamtutut]

Can you give anymore details on what exactly you are going to try to get going with Intel? Are you talking Stratix10 HBM? Or some form of prototype?

Yes, the Stratix 10 parts are equivalent to the Ultrascale+ parts. I believe we'll be able to get the same or better performance for a much lower price point. If FPGA are really going to be viable in the long term to displace GPUs the cost has to be reduced in a significant way. $3995 is ok, right now, while margins are high... But if we really wanted to displace GPUs completely both the cost / perf and power / perf will need to be better than GPUs. I'm hoping that we'll eventually (6-12 months) be able to sell a $1000-2000 board with a stratix 10 or ultrascale+ part that has HBM.

What about this very critical point of view about Intel Stratix 10 HBM ?
https://www.semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-dont-hold-water/

Both Xilinx and Intel are failing to sample or ship HBM chips announced more than six months ago. I don’t think that HBM is coming at mainstream prices in FPGAs anytime soon - even AMD and Nvidia are moving away from it in consumer products.

If we count Vega as consumer cards, it seems AMD is keeping HBM for their next gen, the Vega 20.

[GPUHoarder]

Can you give anymore details on what exactly you are going to try to get going with Intel? Are you talking Stratix10 HBM? Or some form of prototype?

Yes, the Stratix 10 parts are equivalent to the Ultrascale+ parts. I believe we'll be able to get the same or better performance for a much lower price point. If FPGA are really going to be viable in the long term to displace GPUs the cost has to be reduced in a significant way. $3995 is ok, right now, while margins are high... But if we really wanted to displace GPUs completely both the cost / perf and power / perf will need to be better than GPUs. I'm hoping that we'll eventually (6-12 months) be able to sell a $1000-2000 board with a stratix 10 or ultrascale+ part that has HBM.

What about this very critical point of view about Intel Stratix 10 HBM ?
https://www.semiaccurate.com/2017/12/19/intels-claims-fpgas-hbm-dont-hold-water/

Both Xilinx and Intel are failing to sample or ship HBM chips announced more than six months ago. I don’t think that HBM is coming at mainstream prices in FPGAs anytime soon - even AMD and Nvidia are moving away from it in consumer products.

As far as I'm aware they are at least in some limited fashion shipping the stratix 10 mx series. I'll confirm next week for sure.

I think you’re right in that Intel is farther along. They’ve already implemented HBM (or rather MCDRAM) in KNL and have the interposer experience.

I should clarify I don’t think HBM is being abandoned, but it is being relegated to higher cost cards and GDDR6 is taking over in the mass production end. I remember buying 100 Fury X’s in 2015 and having the serial numbers be hand written and I think they incremented maybe a few thousand over the year I accumulated mine.

[Etherion]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

My bad. Sorry mis read the post.

Are you going to start a thread on your m.2 accelerators? Ebay listing?

[GPUHoarder]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

My bad. Sorry mis read the post.

Are you going to start a thread on your m.2 accelerators? Ebay listing?

I’ll start a thread when I’m ready to post all the details - currently expecting that to be about 2-3 weeks.

I’ve got everyone that PM’d me in a list for first dibs, and I’ll be replying with early details soon.

[Etherion]

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.

by then Bitmain FPGA multi algo miner will be available mining away at all the high profile coins.

[senseless]

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.

by then Bitmain FPGA multi algo miner will be available mining away at all the high profile coins.

I would guess this reason is more than likely why Xilinx is hesitant to let anyone else in the crypto community have their chips. Bitmain is already "partnered" with Xilinx.

[trillobeat]

@ GPUHoarder

The M2 accelerator units seem pretty interesting as they would bring some lifespan extension to the GPUs.

You mentioned in a post before that a GPU must be put in 16x or 8x slot and M2 accelerator can boost that  card. Would that mean that boards such as the Asrock B250M Pro4 which have    2x  PCI Express x16 and 2x M2 Ultra slots can accommodate two accelartors for two GPUs plugged in the 16x pci-es?

[Ginzink]

@ GPUHoarder

The M2 accelerator units seem pretty interesting as they would bring some lifespan extension to the GPUs.

You mentioned in a post before that a GPU must be put in 16x or 8x slot and M2 accelerator can boost that  card. Would that mean that boards such as the Asrock B250M Pro4 which have    2x  PCI Express x16 and 2x M2 Ultra slots can accommodate two accelartors for two GPUs plugged in the 16x pci-es?

If having two GPU per MB will be the new way with this mining method, lets prepare for a shortage on motherboards
Quite interesting to see where this leads.

[xaoyao]

I’ve yet to have someone explain to me why they think FPGAs are so bad but GPUs are so good.

Thats a red herring. Who is making that argument ?

The developers who have forked explicitly over their algorithms being listed in this thread.

The economies of scale argument also applies to GPU based mining. For example, with modest capital I could easily build a host system supporting up to 128  GPUs  (since no one uses the PCIe for anything other than making them boot and comm that is lower than serial speed), supporting individually resetting and reinitializing them and all the benefits of smaller hosts, but at 5% of system cost instead of as much as 30%, and also provide power savings. The difference is Bitmain doesn’t represent enough demand to independently control the FPGA market, and the companies in that market are not going to sacrifice decades of high margin business for a short term cash play.

There is no down side for TSMC or commodity memory manufacturers to take big money for cheap part orders from Bitmain or any large player.. There is a downside for Xilinx/Intel flooding the market with cheap FPGAs. Similarly you don’t see NVIDIA and AMD letting Bitmain build custom mining GPUs with their chips at cheap prices. Companies with massive R&D into their chip products want to very carefully maintain control over markets to keep the balance between volume and margin exactly where it is most profitable.

I'm pretty sure intel's plan is to flood the market. Why else would they be developing hybrid cpu/fpgas (like APUs) or why would they bother to create the CCIX interconnect? They're definitely going to flood the market. The question is when. Xilinx will almost surely wait and only react hoping that Intel will also try to preserve high margins. The way I see it, Xilinx has one last chance to gain market share before Intel opens the flood gates. Get your options placed on XLNX while you still can -- Looking forward to making money on their downfall.

Plus, as you and I both know, Quartus is capable of placement / routing that is orders of magnitude better than anything Vivado could do automatically without floorplanning.

I'm looking forward to my call with Intel next week! Hopefully I'll be able to get something rolling quickly and at a lower price point so we (crypto community as a whole) can avoid Xilinx all together.

Edit:

Oh ya, I find it hilarious that devs are changing their algos based on what we say in this thread. Which reminds me, I just developed this new code for cryptonight which makes use of a little hack that's able to bypass some steps to obtain a result quicker. I'm now mining monero at 100Kh/s per fpga.

"I'm now mining monero at 100Kh/s per fpga."
It sounds great. Which FPGA do you use?

[rezin385]

Hes not. He was misleading any coin devs.

[yrk1957]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

Actually from the claims, the Dwarf boards could do 100kh/s at $4000 price-point. If true, ie.

[GPUHoarder]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

Actually from the claims, the Dwarf boards could do 100kh/s at $4000 price-point. If true, ie.

I’d be really interested to see those claims be accurate. I feel like my cryptonight7 is pretty optimal and those numbers are just insane.

[yrk1957]

Whoosh

Are you going to share or just boast?

Read the rest of senseless’ comment - no one actually has 100kh cryptonight on one chip. He was making a commentary on how whole coins are making fork decisions based on comments in this post.

Actually from the claims, the Dwarf boards could do 100kh/s at $4000 price-point. If true, ie.

I’d be really interested to see those claims be accurate. I feel like my cryptonight7 is pretty optimal and those numbers are just insane.

Have you assessed their latest videos?