Mining 2.0 and p106-100 as an affordable GPGPU under float and deep learning
Good day, dear Habrovchane! I’ll tell you a little about mining on video cards and how else you can use the mining graphics cards p106-100 with advantage.
Hayters, pass by. All the rest - if you see a reasonable grain - you can put a plus.
I learned about mining a long time ago, probably in 2015, but I seriously noticed in the middle of 2018 when the GTX 1060 6 Gb video card cost 25,000 each in stores near the city and in general was, as they say now "Big HYIP" on this topic. But even then many smart people predicted not just a recession, but the coming collapse of courses like cryptocurrency and prices for video cards. Even then they predicted that those GTX 1060 would sell for 6,000 miners when the yield drops and it is time to pay loans for which many have collected their farms.
After weighing all the pros and cons then I decided not to risk and not intervene in mining, and I was busy with the project to build the first version of my “village supercomputer”, but I bought a pair of miners blue A006 risers under Molex and a synchronizer. By the way, the synchronizer is used in the server, providing the launch of a kilowatt power supply unit providing power to Tesla K20X cards and their coolers that consume almost 3A each.
Then, on the occasion of the motherboard was bought ASUS H270M-Plus, followed by the cheapest Celeron Q3930 and the cheapest memory to it 2 strips of 4 gigabytes each. Bars 20x40 fir, corners and screws on the tree in my household and so it was. So my first and probably the only rig that stood in the corner was “just in case” where I checked Tesla by loading them with the XMR miner, then I dabbled at mining attempts on the Radeon HD 6450 ... I just stood in the corner and collected dust; there were no maps for him anyway.
')
Actually, everything happened as predicted by smart people. Now the mine card p106-100 6 GB costs about 6-7 tr. apiece and then if samsung memory. And here I actually took two such cards at the beginning of May of 2019 - Gigabyte p106-100 6 Gb Samsung GV-NP106D5-6G rev 1.0. They were taken, by and large, not for mining, but for GPGPU-tasks, as a more economical replacement of Tesla K20 in cases when there is enough float without control of memory errors.
Still, the Tesla K20X has a 235 W TDP and 120 W (with a down-range of 80-83 W) of a p106-100 and a much less noisy cooling system - this is significant during 24/7 operation. And the fact that they do not have video outputs and the output of video cards is forbidden in BIOS did not touch me - cards were not taken for games. Similarly, Tesl has no video outputs, but this completely prevents me from using them for calculations using CUDA.
The cards have arrived. I pulled my rig from the far corner, blew away the dust, picked up the old loyal power supply unit Chiftek 550 W which was a test power supply for the first version of the server, threw the old SSD Kingston 120 Gb, rolled the system with firewood, stuck the cards ... and without becoming wise launched NayshMainer deciding to drive 24/7 weekly cards for burnout defects. It is not enough whether the chip will fall off or the power phase will go out. And he was very surprised when the cards began to bring a couple of euros a day. The very rise of the Bitcoin rate from 4 to 8 thousand dollars happened.
Here is the same rig - “village-stail” with the same pair of Gigabyte p106-100.
But even though I was pleasantly surprised - I ran to take a loan, buy a bag of p106-100, fence a dozen rigs and quit my job hoping to get rich on mining - I did not. Those who have done so already burned out.
But I remembered one interesting article - where the “mining 2.0” concept was mentioned. Its essence is that mining is part of some other process for which the heat generated by the cards is used during operation. For example, a "mining heater" where a large passive radiator was bolted to the four Radeon RX 470 video cards and they were packaged in the form of a heater. Or a project that uses warm air from a working mining farm to heat a greenhouse in the Czech Republic where they grow opiomas. Or other projects of "cryptoboilers" and heating of business premises.
At the time of reading the article, the projects were interesting only as examples of heat utilization of video cards emitted by constantly high loads, the profitability of such a heating system with a price of one card in the region of 20-30 thousand was strictly negative, especially considering the collapse of the cryptocurrency rate that began then. Nevertheless, I rendered useful thoughts from what I read.
These thoughts resulted in an attempt to install on my first Tesla K20M waterblock EKWB from the GTX 780 Ti - which ended in failure. Printed circuit boards 780 and K20 were too different, in the area of the GPU and memory the similarity was still strong, but in the area of the power supply circuits all the elements were placed differently. And as a result, a pair of water blocks on the GTX 780 Ti were stuck with dead weight.
In the current situation, I decided to buy 4 pieces of Gigabyte p106-100 6 Gb by choosing cards on Samsung memory chips that allow some overclocking. Two cards have already been purchased, and I want to purchase 2 more. Since my main purpose of buying these maps is not mining, but GPGPU-tasks, Avito purchased rivers under the GPGPU Tesla model Nvidia Tesla S2050 Server 1U. Of course, I bought him without his “relatives” Tesla, they are too outdated.
Here is a video about this very 1U river.
This device was originally made to connect to any server with a PCI-E 16x 2.0 2-4 Nvidia Tesla M2050 card over a PCI-E 16x 2.0 interface through special cables and Nvidia P797 HIC switch cards. One such switch card connects 2 GPGPU to 1 PCI-E 16x server slot, and the cards work 16x.
The S2050 Rivers contains a power supply that provides cooling fans for the Tesla M2050 server that blows passive radiators, these are the cooling fans, cables with power connectors for 4 PCI-E 6 + 8 pin cards of 225 W each and two cards with a pair of PCI-E 16x slots , switch chips under passive radiators and connectors for communication with the server. Despite the fact that the rivers are proudly named the “Nvidia Tesla S2050 server” - it is just a 4 GPGPU box with fans, a power supply unit and risers.
Actually when I have 4 Gigabyte p106-100 cards, and cables and Nvidia P797 HIC cards will come to me, the cards will be installed in the rivers, the rivers will be connected to the remaining free first server board and I will see what happens. Still, 4 GPGPUs, even p106-100, but connected via a full-fledged PCI-E 16x 2.0 bus — in tasks where there is enough float and no memory error monitoring is required — very good. And first of all, this is work with frameworks implementing software neural networks, the very deep learning. Unlike classical engineering and other computations, neural network algorithms are resistant to “equipment errors” and the loss of individual memory bits on GTX video games or mining p106-100 will not cause problems. But the p106-100 is much more affordable than the Tesla K20 and consumes noticeably less energy.
And now a little about Gigabyte P106-100 GV-NP106D5-6G rev 1.0 themselves. Of course, I could not help but make out the cards - it was interesting for me to look at the board without radiators and find out from which model of the “ordinary” GTX 1060 itself this “miracle” was made? Interest was not only research. I wanted to find water blocks on their “direct ancestors” - because you have not forgotten about the “mining heating”? Of course, I do not plan to engage in mining - but I still want to divert the heat produced without a howling of an airplane taking off and it is advisable to use it with benefit. And here the most convenient option is a liquid cooling system.
A Phillips screwdriver, straight arms, and here it is - a scan of my p106-100 board on both sides.
Video chip side.
The reverse side of the board.
Comparison of my Gigabyte P106-100 GV-NP106D5-6G rev 1.0 board with photo cards of all kinds of GTX 1060 6 Gb with similar and not very fans brought their results. "Next of kin" were installed.
And already knowing the “blood kin” model all on the same Ali, a link was found to the water block that matched my p106-100, the Chinese confirmed the compatibility by scan.
Here is a link to it.
Now, in fact, you know everything you need to assemble a mining heater on these p106-100.
In my plans, first of all, to bring the number of my Gigabyte p106-100 to 4, and then, as far as possible, to 8. The second river Nvidia Tesla S2050 is also bought by me and the transport company is going to me. The second stage, of course, when the system will already work as a GPGPU, is planned to transfer the cooling system to ITS using those same Chinese water blocks with Ali and further practical use of the heat generated by the system for household needs.
In the process of assembling the p106-100 card system, they will be checked by me under a long load in that mining, and if they can pay back part of the cost of the system being assembled, so much the better. If not, then at least they would be checked under a constantly high load and the risk of a card failure in an already assembled system will be noticeably reduced.
In addition, the assembled system will be used by me mainly for deep learning and maybe I will also participate in the Boinc distributed computing project where participants provide their computing resources for scientific calculations "For the good of all mankind."
To whom the article seemed interesting - subscribe to my channel on YouTube, there is the latest information and regularly posted new interesting videos.
Hayters, pass by. All the rest - if you see a reasonable grain - you can put a plus.
I learned about mining a long time ago, probably in 2015, but I seriously noticed in the middle of 2018 when the GTX 1060 6 Gb video card cost 25,000 each in stores near the city and in general was, as they say now "Big HYIP" on this topic. But even then many smart people predicted not just a recession, but the coming collapse of courses like cryptocurrency and prices for video cards. Even then they predicted that those GTX 1060 would sell for 6,000 miners when the yield drops and it is time to pay loans for which many have collected their farms.
After weighing all the pros and cons then I decided not to risk and not intervene in mining, and I was busy with the project to build the first version of my “village supercomputer”, but I bought a pair of miners blue A006 risers under Molex and a synchronizer. By the way, the synchronizer is used in the server, providing the launch of a kilowatt power supply unit providing power to Tesla K20X cards and their coolers that consume almost 3A each.
Then, on the occasion of the motherboard was bought ASUS H270M-Plus, followed by the cheapest Celeron Q3930 and the cheapest memory to it 2 strips of 4 gigabytes each. Bars 20x40 fir, corners and screws on the tree in my household and so it was. So my first and probably the only rig that stood in the corner was “just in case” where I checked Tesla by loading them with the XMR miner, then I dabbled at mining attempts on the Radeon HD 6450 ... I just stood in the corner and collected dust; there were no maps for him anyway.
')
Actually, everything happened as predicted by smart people. Now the mine card p106-100 6 GB costs about 6-7 tr. apiece and then if samsung memory. And here I actually took two such cards at the beginning of May of 2019 - Gigabyte p106-100 6 Gb Samsung GV-NP106D5-6G rev 1.0. They were taken, by and large, not for mining, but for GPGPU-tasks, as a more economical replacement of Tesla K20 in cases when there is enough float without control of memory errors.
Still, the Tesla K20X has a 235 W TDP and 120 W (with a down-range of 80-83 W) of a p106-100 and a much less noisy cooling system - this is significant during 24/7 operation. And the fact that they do not have video outputs and the output of video cards is forbidden in BIOS did not touch me - cards were not taken for games. Similarly, Tesl has no video outputs, but this completely prevents me from using them for calculations using CUDA.
The cards have arrived. I pulled my rig from the far corner, blew away the dust, picked up the old loyal power supply unit Chiftek 550 W which was a test power supply for the first version of the server, threw the old SSD Kingston 120 Gb, rolled the system with firewood, stuck the cards ... and without becoming wise launched NayshMainer deciding to drive 24/7 weekly cards for burnout defects. It is not enough whether the chip will fall off or the power phase will go out. And he was very surprised when the cards began to bring a couple of euros a day. The very rise of the Bitcoin rate from 4 to 8 thousand dollars happened.
Here is the same rig - “village-stail” with the same pair of Gigabyte p106-100.
But even though I was pleasantly surprised - I ran to take a loan, buy a bag of p106-100, fence a dozen rigs and quit my job hoping to get rich on mining - I did not. Those who have done so already burned out.
But I remembered one interesting article - where the “mining 2.0” concept was mentioned. Its essence is that mining is part of some other process for which the heat generated by the cards is used during operation. For example, a "mining heater" where a large passive radiator was bolted to the four Radeon RX 470 video cards and they were packaged in the form of a heater. Or a project that uses warm air from a working mining farm to heat a greenhouse in the Czech Republic where they grow opiomas. Or other projects of "cryptoboilers" and heating of business premises.
At the time of reading the article, the projects were interesting only as examples of heat utilization of video cards emitted by constantly high loads, the profitability of such a heating system with a price of one card in the region of 20-30 thousand was strictly negative, especially considering the collapse of the cryptocurrency rate that began then. Nevertheless, I rendered useful thoughts from what I read.
These thoughts resulted in an attempt to install on my first Tesla K20M waterblock EKWB from the GTX 780 Ti - which ended in failure. Printed circuit boards 780 and K20 were too different, in the area of the GPU and memory the similarity was still strong, but in the area of the power supply circuits all the elements were placed differently. And as a result, a pair of water blocks on the GTX 780 Ti were stuck with dead weight.
In the current situation, I decided to buy 4 pieces of Gigabyte p106-100 6 Gb by choosing cards on Samsung memory chips that allow some overclocking. Two cards have already been purchased, and I want to purchase 2 more. Since my main purpose of buying these maps is not mining, but GPGPU-tasks, Avito purchased rivers under the GPGPU Tesla model Nvidia Tesla S2050 Server 1U. Of course, I bought him without his “relatives” Tesla, they are too outdated.
Here is a video about this very 1U river.
This device was originally made to connect to any server with a PCI-E 16x 2.0 2-4 Nvidia Tesla M2050 card over a PCI-E 16x 2.0 interface through special cables and Nvidia P797 HIC switch cards. One such switch card connects 2 GPGPU to 1 PCI-E 16x server slot, and the cards work 16x.
The S2050 Rivers contains a power supply that provides cooling fans for the Tesla M2050 server that blows passive radiators, these are the cooling fans, cables with power connectors for 4 PCI-E 6 + 8 pin cards of 225 W each and two cards with a pair of PCI-E 16x slots , switch chips under passive radiators and connectors for communication with the server. Despite the fact that the rivers are proudly named the “Nvidia Tesla S2050 server” - it is just a 4 GPGPU box with fans, a power supply unit and risers.
Actually when I have 4 Gigabyte p106-100 cards, and cables and Nvidia P797 HIC cards will come to me, the cards will be installed in the rivers, the rivers will be connected to the remaining free first server board and I will see what happens. Still, 4 GPGPUs, even p106-100, but connected via a full-fledged PCI-E 16x 2.0 bus — in tasks where there is enough float and no memory error monitoring is required — very good. And first of all, this is work with frameworks implementing software neural networks, the very deep learning. Unlike classical engineering and other computations, neural network algorithms are resistant to “equipment errors” and the loss of individual memory bits on GTX video games or mining p106-100 will not cause problems. But the p106-100 is much more affordable than the Tesla K20 and consumes noticeably less energy.
And now a little about Gigabyte P106-100 GV-NP106D5-6G rev 1.0 themselves. Of course, I could not help but make out the cards - it was interesting for me to look at the board without radiators and find out from which model of the “ordinary” GTX 1060 itself this “miracle” was made? Interest was not only research. I wanted to find water blocks on their “direct ancestors” - because you have not forgotten about the “mining heating”? Of course, I do not plan to engage in mining - but I still want to divert the heat produced without a howling of an airplane taking off and it is advisable to use it with benefit. And here the most convenient option is a liquid cooling system.
A Phillips screwdriver, straight arms, and here it is - a scan of my p106-100 board on both sides.
Video chip side.
The reverse side of the board.
Comparison of my Gigabyte P106-100 GV-NP106D5-6G rev 1.0 board with photo cards of all kinds of GTX 1060 6 Gb with similar and not very fans brought their results. "Next of kin" were installed.
- Gigabyte GTX 1060 WindForce OC 6G Gigabyte GV-N1060WF2OC-6GD Link
- Gigabyte GTX 1060 Mini ITX OC 3G Link
And already knowing the “blood kin” model all on the same Ali, a link was found to the water block that matched my p106-100, the Chinese confirmed the compatibility by scan.
Here is a link to it.
Now, in fact, you know everything you need to assemble a mining heater on these p106-100.
In my plans, first of all, to bring the number of my Gigabyte p106-100 to 4, and then, as far as possible, to 8. The second river Nvidia Tesla S2050 is also bought by me and the transport company is going to me. The second stage, of course, when the system will already work as a GPGPU, is planned to transfer the cooling system to ITS using those same Chinese water blocks with Ali and further practical use of the heat generated by the system for household needs.
In the process of assembling the p106-100 card system, they will be checked by me under a long load in that mining, and if they can pay back part of the cost of the system being assembled, so much the better. If not, then at least they would be checked under a constantly high load and the risk of a card failure in an already assembled system will be noticeably reduced.
In addition, the assembled system will be used by me mainly for deep learning and maybe I will also participate in the Boinc distributed computing project where participants provide their computing resources for scientific calculations "For the good of all mankind."
To whom the article seemed interesting - subscribe to my channel on YouTube, there is the latest information and regularly posted new interesting videos.
Source: https://habr.com/ru/post/454560/
All Articles