Topic: DIY FPGA Mining rig for any algorithm with fast ROI - page 92. (Read 99472 times)
where is fpga-claymore
VCU1525 has 2 x QSFP28 connectors, so you link 2 boards with 2 x 100 gigabit ethernet cables, and two boards is the max that can be daisy chained. With the Bittware XUPP3R board, since it has 4 x QSFP28 connectors, there is no limit to the daisy chain length. These specialized ethernet cables have nothing to do with internet, they are solely so data can flow from one FPGA board to the next board. The cables are around $40 each.
I was wondering about other ways of setting up the communication between the boards.Looking at the VCU1525 user's guide they have 8 transceivers hooked to Ethernet ports and 16 transceivers hooked to PCIe, remaining transceiver blocks aren't connected. Also, it looks like the required intellectual property block for the PCIe support.
Have you considered plugging those boards into a PCIe passive backplane type of interface, not to the actual PCIe computer motherboard? This should be very simple to interface, even to the point of not really obeying the official PCIe protocol rules if the passive backplane contains only the boards running our project.
Can you write a paragraph or two about the relative logic overhead required to implement inter-board communication over Ethernet versus PCIe?
I currently have all my development hardware in storage, so I can't even make a dry-run with an evaluation version of the software.
what you need is that : https://www.youtube.com/watch?time_continue=2&v=QNMKh2lC_60
For Ravencoin and Bitcore, the DDR4 is not used for hashing at all, but rather to store hundreds of different FPGA configuration bitstreams, which allows the entire FPGA to rapidly reprogram itself on every block based on the algorithm sequence for that block. Cryptonight7 and Equihash, on the other hand, would use the external memory for actual hashing. Whether CN7/Equihash gain from additional RAM depends on how those algorithms are ultimately implemented.
Hey, we have opened a discord chat to avoid flooding this thread, we'll be glad to have you there (70 people at the moment)
https://discord.gg/CTgdmy
I know you are busy but it could help all parties...
For Ravencoin and Bitcore, the DDR4 is not used for hashing at all, but rather to store hundreds of different FPGA configuration bitstreams, which allows the entire FPGA to rapidly reprogram itself on every block based on the algorithm sequence for that block. Cryptonight7 and Equihash, on the other hand, would use the external memory for actual hashing. Whether CN7/Equihash gain from additional RAM depends on how those algorithms are ultimately implemented.
Does anyone know how it works if you purchase the Bittware version since the defaults is w/no memory installed? Can you purchase your own DDR memory to install, and if so what kinda?
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
My Bittware board has no external memory, and I will add either standard PC DDR4, or Bittware QDR-SRAM, as needed in the future; only Ravencoin (DDR4), Bitcore (DDR4) and Equihash (QDR) need external memory. Cryptonight7 would, although I haven't examined it in depth. Phi, Keccak, SHA-224, Tribus, X17, Lyra2v2, Neoscrypt, Xevan; no external memory needed.
Will the hash increase be equal to the amount of external memory added? For ex. will going from 16GB to 32GB double the hash etc..? Also, is QDR specific to Equihash or is also compatible with RVN/Bitcore?
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
Yes of course. I thought LTCMAXMYR compares 1080ti with 28 nm FPGA.
keccak 2 core ,@200MHz，may run 300Mhz,but Limited by chip specification and board 1.0V DC/DC power.
28nm kintex 325t, 400M, 25W , ~2.1A @12V, so 16M/W
same time GeForceGTX970 got 400M @125W , 3.2M/W
xilinx report 16nm 2x improvment
https://www.xilinx.com/products/technology/power.html
16nm XCVU9P may get 32M/W
16nm GTX1080ti ,1200M, 150W ，8M/W
that is about 4x
14nm GPU will be more efficient，so it can be lower than 4x
I dont understand.You says OP says 650MH on phi with 150W which is 4.3Mh/W
1080ti makes 30mh with 150W which is 0.2Mh/W which is x20 more efficient.
which is it x4 times more or x20 times more efficient?
so yours calculation x4 times more efficient than 1080ti is applicable only for keccak?
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
Yes of course. I thought LTCMAXMYR compares 1080ti with 28 nm FPGA.
keccak 2 core ,@200MHz，may run 300Mhz,but Limited by chip specification and board 1.0V DC/DC power.
28nm kintex 325t, 400M, 25W , ~2.1A @12V, so 16M/W
same time GeForceGTX970 got 400M @125W , 3.2M/W
xilinx report 16nm 2x improvment
https://www.xilinx.com/products/technology/power.html
16nm XCVU9P may get 32M/W
16nm GTX1080ti ,1200M, 150W ，8M/W
that is about 4x
14nm GPU will be more efficient，so it can be lower than 4x
I dont understand.You says OP says 650MH on phi with 150W which is 4.3Mh/W
1080ti makes 30mh with 150W which is 0.2Mh/W which is x20 more efficient.
which is it x4 times more or x20 times more efficient?
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
Yes of course. I thought LTCMAXMYR compares 1080ti with 28 nm FPGA.
keccak 2 core ,@200MHz，may run 300Mhz,but Limited by chip specification and board 1.0V DC/DC power.
28nm kintex 325t, 400M, 25W , ~2.1A @12V, so 16M/W
same time GeForceGTX970 got 400M @125W , 3.2M/W
xilinx report 16nm 2x improvment
https://www.xilinx.com/products/technology/power.html
16nm XCVU9P may get 32M/W
16nm GTX1080ti ,1200M, 150W ，8M/W
that is about 4x
14nm GPU will be more efficient，so it can be lower than 4x
I dont understand.You says OP says 650MH on phi with 150W which is 4.3Mh/W
1080ti makes 30mh with 150W which is 0.2Mh/W which is x20 more efficient.
which is it x4 times more or x20 times more efficient?
Does anyone know how it works if you purchase the Bittware version since the defaults is w/no memory installed? Can you purchase your own DDR memory to install, and if so what kinda?
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
My Bittware board has no external memory, and I will add either standard PC DDR4, or Bittware QDR-SRAM, as needed in the future; only Ravencoin (DDR4), Bitcore (DDR4) and Equihash (QDR) need external memory. Cryptonight7 would, although I haven't examined it in depth. Phi, Keccak, SHA-224, Tribus, X17, Lyra2v2, Neoscrypt, Xevan; no external memory needed.
Yes, you are right in this statement, but I do not want to write a long story. It seems that Ethash is only algorithm that is not suitable for FPGA.
Not so fast, there is a guy who claimed he can do Ethash at 421.41 MH (FPGA Nallatech 385A). I am still waiting for proof, so it may be possible. Or I may be waiting for a very long time.
https://github.com/Maetti79/OpenCL_FPGA
https://github.com/Maetti79/ethminer
Thanks for the links. According to his latest results, the FPGA takes ETH hash from MEGA to GIGA.
Does anyone know how it works if you purchase the Bittware version since the defaults is w/no memory installed? Can you purchase your own DDR memory to install, and if so what kinda?
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
The following is offered from Bittware, but what is recommended?
R4 = DDR4 16GB RDIMM
R5 = DDR4 32GB RDIMM
R7 = DDR4 128GB RDIMM
L5 = DDR4 32GB LRDIMM
L6 = DDR4 64GB LRDIMM
Q4 = Dual QDRII+ x18 144Mb
Q5 = Dual QDRII+ x18 288Mb
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
Yes of course. I thought LTCMAXMYR compares 1080ti with 28 nm FPGA.
keccak 2 core ,@200MHz，may run 300Mhz,but Limited by chip specification and board 1.0V DC/DC power.
28nm kintex 325t, 400M, 25W , ~2.1A @12V, so 16M/W
same time GeForceGTX970 got 400M @125W , 3.2M/W
xilinx report 16nm 2x improvment
https://www.xilinx.com/products/technology/power.html
16nm XCVU9P may get 32M/W
16nm GTX1080ti ,1200M, 150W ，8M/W
that is about 4x
14nm GPU will be more efficient，so it can be lower than 4x
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
Yes of course. I thought LTCMAXMYR compares 1080ti with 28 nm FPGA.
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
The VU9P FPGA is a 16nm device, not 28nm.
I admit that I'm afraid of this device. I just set my rig up late December before the peak and the crash.
The last thing I want to do is see this come out and totally annihilate GPU mining. And that is exactly what it has the potential to do. Profitability for everything else will be completely destroyed and we'll be stuck fire selling everything we have in the hope we can get in line to acquire this hardware and software. That is if the promises of being able to actually obtain this in the first place without large farms buying up everything and dominating are true.
That metroid kid may be right. This could be the end of GPU mining.
The last thing I want to do is see this come out and totally annihilate GPU mining. And that is exactly what it has the potential to do. Profitability for everything else will be completely destroyed and we'll be stuck fire selling everything we have in the hope we can get in line to acquire this hardware and software. That is if the promises of being able to actually obtain this in the first place without large farms buying up everything and dominating are true.
That metroid kid may be right. This could be the end of GPU mining.
do not be afraid.
1:1 perf FPGA VS 1080TI
POWER, 1:4 ，
PRICE , 10:1
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
These estimates are closer to reality. Do not panic, but watch your profit.
2:1 perf mid range FPGA VS 1080TI (* when algorithm is supported)
POWER, 1:4...5 (** ok)
PRICE , 1:1 (*** in mass production)
VCU1525 has 2 x QSFP28 connectors, so you link 2 boards with 2 x 100 gigabit ethernet cables, and two boards is the max that can be daisy chained. With the Bittware XUPP3R board, since it has 4 x QSFP28 connectors, there is no limit to the daisy chain length. These specialized ethernet cables have nothing to do with internet, they are solely so data can flow from one FPGA board to the next board. The cables are around $40 each.
I was wondering about other ways of setting up the communication between the boards.Looking at the VCU1525 user's guide they have 8 transceivers hooked to Ethernet ports and 16 transceivers hooked to PCIe, remaining transceiver blocks aren't connected. Also, it looks like the required intellectual property block for the PCIe support.
Have you considered plugging those boards into a PCIe passive backplane type of interface, not to the actual PCIe computer motherboard? This should be very simple to interface, even to the point of not really obeying the official PCIe protocol rules if the passive backplane contains only the boards running our project.
Can you write a paragraph or two about the relative logic overhead required to implement inter-board communication over Ethernet versus PCIe?
I currently have all my development hardware in storage, so I can't even make a dry-run with an evaluation version of the software.
VCU1525 has 2 x QSFP28 connectors, so you link 2 boards with 2 x 100 gigabit ethernet cables, and two boards is the max that can be daisy chained. With the Bittware XUPP3R board, since it has 4 x QSFP28 connectors, there is no limit to the daisy chain length. These specialized ethernet cables have nothing to do with internet, they are solely so data can flow from one FPGA board to the next board. The cables are around $40 each.
Personally I think GPU mining will still be around for a while. I believe the market is going to rise so dramatically this fall that all existing GPU rigs will be making a lot of money, even with rises in difficulty.
Personally I think GPU mining will still be around for a while. I believe the market is going to rise so dramatically this fall that all existing GPU rigs will be making a lot of money, even with rises in difficulty.
Whitefire are you willing to meet in person to demonstrate your work? You have a PM about it. Please let me know.
With x16r and x17 requiring 2 cards would that be 300mh/s for both cards? Or 300 each equalling 600mh for two cards daisy chained together?
Clarifying the projected hash rates
X17: 2 cards daisy chained get 600MH/s total
X16R: 2 cards daisy chained get 600MH/s total
Xevan: 4 Bittware cards daisy chained get 600MH/s total
VCU1525 has 2 x QSFP28 connectors, so you link 2 boards with 2 x 100 gigabit ethernet cables, and two boards is the max that can be daisy chained. With the Bittware XUPP3R board, since it has 4 x QSFP28 connectors, there is no limit to the daisy chain length. These specialized ethernet cables have nothing to do with internet, they are solely so data can flow from one FPGA board to the next board. The cables are around $40 each.
Personally I think GPU mining will still be around for a while. I believe the market is going to rise so dramatically this fall that all existing GPU rigs will be making a lot of money, even with rises in difficulty.
If this doomsday device is what the OP suggests. I don't see how GPUs are going to survive this. If the ASIC is a 10 kiloton hiroshima style weapon. This is a 10 megaton hydrogen bomb. And it's going to get dropped right on your GPU rig.
I wet the bed last night after reading this thread.
I wet the bed last night after reading this thread.
I admit that I'm afraid of this device. I just set my rig up late December before the peak and the crash.
The last thing I want to do is see this come out and totally annihilate GPU mining. And that is exactly what it has the potential to do. Profitability for everything else will be completely destroyed and we'll be stuck fire selling everything we have in the hope we can get in line to acquire this hardware and software. That is if the promises of being able to actually obtain this in the first place without large farms buying up everything and dominating are true.
That metroid kid may be right. This could be the end of GPU mining.
The last thing I want to do is see this come out and totally annihilate GPU mining. And that is exactly what it has the potential to do. Profitability for everything else will be completely destroyed and we'll be stuck fire selling everything we have in the hope we can get in line to acquire this hardware and software. That is if the promises of being able to actually obtain this in the first place without large farms buying up everything and dominating are true.
That metroid kid may be right. This could be the end of GPU mining.
do not be afraid.
1:1 perf FPGA VS 1080TI
POWER, 1:4 ，
PRICE , 10:1
I doubt that FPGA performance/perwatt,same 28nm, FPGA can only get 5x lower power，14nm GPU is more efficient，so it can be 4x
Jump to: