DIY bitcoin miner for Avalon A3255-Q48 chips

[FrankBuss]

With 10 avalon chips per module, what's the plan if one chip fails? Will the chip be replaced or the entire module? How easy will it be to swap modules?

Replacing a chip could be difficult, because you have to desolder the QFN48 package, clean it and then solder a new chip. Maybe I should add some jumpers or 0-ohm resistors, then a broken chip could be disconnected easier. And some more non-populated 0-ohm bridges to skip the CONFIG_IN/CONFIG_OUT chain for a broken chip. It should be possible to detect a faulty chip by software, at least the first one in a chain, if multiple are broken.

Here's my very high-level understanding of your rig so far:

1,000 avalon chips, with 10 chips per module for 100 modules.

Will it be stacked, spread out flat or side-by-side or ?

What about fire supression?

Good questions, I have no idea I'm a programmer and can do some electronics, but no mechanics. I guess it would end up looking like this:



Meanwhile I've done some more tests. When I run my test setup with 1 GH/s, the report signal has a frequency of 4 MHz. This would be difficult to sample with a microcontroller, especially if I want to do something more in parallel, like measuring the temperature and communicating with the controller board. But I developed a circuit to convert the Avalon REPORT_N/REPORT_P signal to a SPI signal:



And this is how it looks like with real signals:



As you can see, the OR_DELAY signal can be used as the SPI clock (configured to sample MOSI at rising edge) and RPT_N_I as MOSI. Two 30 cent chips replaces the whole FPGA! And I tested the configure signals as well: there is no lower limit how slow it can be, so no need for the 125 ns signal generating part of the FPGA either.

This would allow a cheap PIC for each Avalon module. But I think I'll use a more expensive ARM microcontroller with USB (4 EUR instead of 2 EUR). Then I can design the 10 chip Avalon module to run in standalone mode with USB, or connected over SPI to the Raspberry Pi controller board. And the microcontroller can be programmed over USB as well, no need for a programmer, which makes it easier for others to rebuild it.

[GiskardReventlov]

Replacing a chip could be difficult, because you have to desolder the QFN48 package, clean it and then solder a new chip.

This would work but wow they are expensive.

http://www.mouser.com/access/?pn=945-IC550-0484-004-G

As you can see, the OR_DELAY signal can be used as the SPI clock (configured to sample MOSI at rising edge) and RPT_N_I as MOSI. Two 30 cent chips replaces the whole FPGA! And I tested the configure signals as well: there is no lower limit how slow it can be, so no need for the 125 ns signal generating part of the FPGA either.

Since I have to look up what MOSI is you can guess I have nothing to say about this.

This would allow a cheap PIC for each Avalon module. But I think I'll use a more expensive ARM microcontroller with USB (4 EUR instead of 2 EUR). Then I can design the 10 chip Avalon module to run in standalone mode with USB, or connected over SPI to the Raspberry Pi controller board. And the microcontroller can be programmed over USB as well, no need for a programmer, which makes it easier for others to rebuild it.

After looking at the pic with all the wires maybe instead of usb go wireless?  Bluetooth or wifi?  But I'm curious about the ARM microcontroller w/ USB. What part is that?

[scientist]

Why not bury your server in the permafrost and use a Peltier plate to generate power from the surplus heat from the server? Then, power the fans with it.

[Bassman59]

Since I have to look up what MOSI is you can guess I have nothing to say about this.

It's one of the SPI data signals, "Master Out, Slave In."

-a

[FrankBuss]

This would work but wow they are expensive.

http://www.mouser.com/access/?pn=945-IC550-0484-004-G

Yes, such sockets are expensive. Usually it is used for programming or testing chips. And if I would use it, there would be a problem with the exposed center pad, which has to be soldered for best heat transfer. But could be used for a Avalon chip tester. Maybe a good idea to test all chips before soldering.

After looking at the pic with all the wires maybe instead of usb go wireless?  Bluetooth or wifi?  But I'm curious about the ARM microcontroller w/ USB. What part is that?

My idea was a LPC11U2X controller, as I used for my Crazy Cartridge. But unfortunately the fine print in the datasheet says the maximum SPI clock is 1/12 system clock. For the 4 MHz SPI clock I would need a system clock of 48 MHz. This is exactly the frequency with which it is running, so there would be no safety margin and if I overclock the Avalon, it could fail. And the SPI slave implementation on the LPC chip is not really usable: it needs a CS signal to start the receiving, and max word size is 16 bit. I would need to use another pin as an output pin to simulate CS and I would need to implement some interrupts to read two words.

So back to a PIC. The PIC32MX220F032B looks perfect. The SPI slave implementation allows a SPI clock up to 20 MHz, it can be configured for 32 bit word size and it receives without the need for a CS signal (but which can be enabled). And it is cheap.

I think for the communication from the controller to the Avalon modules I would use a much simpler concept, a chained UART line: The controller board has an UART tx line, connected to RX of the PIC on the first board. The PIC has a TX line, connected to the RX line on the second board etc. and the tx line of the last board connected back to the RX line of the controller board. This would solve the cable mess, and it could be connected to a (TTL) serial port of a PC as well, without any controller board, which could be simply a Raspberry Pi, without any custom PCB.

[GiskardReventlov]

And if I would use it, there would be a problem with the exposed center pad, which has to be soldered for best heat transfer. But could be used for a Avalon chip tester. Maybe a good idea to test all chips before soldering.

I didn't think about the center pad, but yes it would be good  for pre-solder test.

So back to a PIC. The PIC32MX220F032B looks perfect. The SPI slave implementation allows a SPI clock up to 20 MHz, it can be configured for 32 bit word size and it receives without the need for a CS signal (but which can be enabled). And it is cheap.

I think for the communication from the controller to the Avalon modules I would use a much simpler concept, a chained UART line: The controller board has an UART tx line, connected to RX of the PIC on the first board. The PIC has a TX line, connected to the RX line on the second board etc. and the tx line of the last board connected back to the RX line of the controller board. This would solve the cable mess, and it could be connected to a (TTL) serial port of a PC as well, without any controller board, which could be simply a Raspberry Pi, without any custom PCB.

I have never done any work with a PIC.  I'm going to look into that. I need to take the cheapest route though.

Here's my understanding, each Avalon work's on  it's own sha256(sha256()) and will return data at a rate distinct from every other Avalon. So the data rates from 1000 Avalons should be expected to be small. Or am I missing something? i.e. the Avalon spends more time working than talking.

[FrankBuss]

Shouldn't be too expensive. A PICKit 3 programmer clone for the PIC32 parts costs EUR 20 at eBay and one PIC32MX220F032B costs EUR 2.43.

You are right with the low data rate. From my tests I see 0 to 3 nonces for 4*10^9 tests. At the highest rated speed of 1 GH/s this would result in maybe one nonce per chip, so for the full version with 1,000 chips it would be 1,000 nonces per second. It needs 32 bits to transfer it, and maybe another byte to tell the serial chain that it is a nonce report, which would result in 40 kHz. Should be possible with a 115,200 baud serial port.

But the PICs have to do some more calculation to keep the Avalons running, because 1,000 Avalons can't be updated fast enough over the serial line. If I implement the getblocktemplate concept in the PIC, they can mine a long time for one getblocktemplate response from a pool, without the need to be configured from the main controller all the time. The PIC32 is powerful enough for this algorithm.

[GiskardReventlov]

Shouldn't be too expensive. A PICKit 3 programmer clone for the PIC32 parts costs EUR 20 at eBay and one PIC32MX220F032B costs EUR 2.43.

I did find them, they are cheap.

You are right with the low data rate. From my tests I see 0 to 3 nonces for 4*10^9 tests. At the highest rated speed of 1 GH/s this would result in maybe one nonce per chip, so for the full version with 1,000 chips it would be 1,000 nonces per second. It needs 32 bits to transfer it, and maybe another byte to tell the serial chain that it is a nonce report, which would result in 40 kHz. Should be possible with a 115,200 baud serial port.

Ok  1000*5 bytes = 5000 bytes/sec but I didn't follow you on the 40 kHz.

But the PICs have to do some more calculation to keep the Avalons running, because 1,000 Avalons can't be updated fast enough over the serial line. If I implement the getblocktemplate concept in the PIC, they can mine a long time for one getblocktemplate response from a pool, without the need to be configured from the main controller all the time. The PIC32 is powerful enough for this algorithm.

There's one PIC per module (or one PIC for every 10 Avalons)? Or can you spread out the PIC to handle more than one module?

[FrankBuss]

5,000 bytes per second means 40,000 bits per second, so it would need a transfer bandwidth of 40 kHz (ignoring start and stop bits). That's lower than 115,200 baud of a standard serial port, so it should work, even with a PC as the controller. Might need some buffering on the PICs in the chain, but they have plenty of RAM.

I was planning one PIC per module. The Avalon datasheet doesn't say how many Avalons can be chained, the signal may deteriorate with too chips. And it makes the design easier: one PIC for one module, with one temperature sensor per PIC and one Avalon chain.

I think there are boards with 20 Avalons in chain, but there might be other problems, because it would need 0.9 V @ 40 A, which requires big traces. With 10 Avalons probably I can use a cheaper 2 layer board per module.

[GiskardReventlov]

5,000 bytes per second means 40,000 bits per second, so it would need a transfer bandwidth of 40 kHz (ignoring start and stop bits). That's lower than 115,200 baud of a standard serial port, so it should work, even with a PC as the controller. Might need some buffering on the PICs in the chain, but they have plenty of RAM.

Doesn't seem like an unrealistic setup.  Those rigs with all the USB cables were way over-engineered. But were those ASICs or FPGA miners?

I was planning one PIC per module. The Avalon datasheet doesn't say how many Avalons can be chained, the signal may deteriorate with too chips. And it makes the design easier: one PIC for one module, with one temperature sensor per PIC and one Avalon chain.

I think there are boards with 20 Avalons in chain, but there might be other problems, because it would need 0.9 V @ 40 A, which requires big traces. With 10 Avalons probably I can use a cheaper 2 layer board per module.

Did you already say what temp. sensor?  What's cost per module at now?
I see
1 avalon * $12 = 10 per module = $120
1 PIC = $3
Temp sensors = ?
2-layer PCB = ?
serial connectors = ?
incidentals = solder,flux,etc. = nex to nil
labor  = fun = free = ?
 

ebay international lists lots of avalon chips but most in the U.S. seem to be plug-n-play mining rigs. I did see  single lot of chips but they were much more expensive compared to the international listings.  $5-6 per compared to $18 per.

[FrankBuss]

This is a first version of the schematic with the new PIC: https://github.com/FrankBuss/gandalf-minder/raw/master/eagle/miner.pdf
Some planned parts with Digikey part numbers:

CPU: PIC32MX220F032B-I/SO-ND
OR gate: 74LVC1G14SE-7CT-ND
hex schmitt trigger: 74LCX14MXCT-ND
25 MHz oscillator: 336-2942-ND
temperature sensor: TCN75AVUA-ND

But the most expensive part might be the 0.9 V voltage regulator with at least 20 A output (and a digital potentiometer to adjust the voltage). I found a nice module, but it costs EUR 17.84. I have no experience with designing power supplies, so I might use this module for a first version, and in a later version I can try to integrate the individual parts on the board.

The serial connector from board to board and to the Raspberry Pi could be just pin headers and jumper wires.

And it would need a good heatsink. I can't find a good one at Digikey, but this one looks good:
http://de.farnell.com/jsp/search/productdetail.jsp?SKU=4621980
EUR 16.34 for a nice big 20 cm x 10 cm heatsink.
And finally one PCB with 25 cm x 10 cm costs $13.60 at Iteadstudio (10 cm x 20 cm for the Avalon area and 5 cm x 10 cm for the PIC and the rest).

In sum it might cost EUR 50 for one module, without the Avalons. This is half the price of for example this board:
http://www.burninmining.com/product-category/boards/
It has 20 Avalons, maybe this would be a good idea for my board as well.

[nctnico]

Toploser on this forum still has some nice heatsinks AFAIK (look in the buy/sell/wanted section). You could mount several modules on them. These heatsinks need a fan to be effective though.

Looking at the bitburner I'm not sure whether the heatsink should be placed on the solder side of the PCB. It seems more logical to place the heatsink on top of the chips (with heat conductive rubber in between). Perhaps others can chime in on this subject.

[FrankBuss]

Good idea with the heatsinks, bigger than what I found and cheaper. I'll ask him if he still has some left for sale.

Placing the heatsink on the back is better for QFN parts with exposed center pads, because the combined thermal resistance is lower than placing the heatsink on top of the chips (see for example this table: http://cds.linear.com/docs/en/packaging/Linear_Technology_Thermal_Resistance_Table.pdf ). Of course, needs good heat conduction to the back side with some vias in the center pad.

[tom66]

You are using a linear regulator for 0.9V core voltage?

[GiskardReventlov]

In sum it might cost EUR 50 for one module, without the Avalons. This is half the price of for example this board:
http://www.burninmining.com/product-category/boards/
It has 20 Avalons, maybe this would be a good idea for my board as well.

That board says 5.6GH/s per board.  They are using different chips?

Would there be any advantage to keeping the avalon board separate from the PIC and the rest? I was thinking that having a 10-avalon board with just the serial and power. Then have the PIC and the rest on another board. This may make it possible to swap out the 10-avalon board easily and possibly give an upgrade path for the next-gen avalon.  Maybe make it more expensive but possibly offer reuse or repurpose even.


I'm not clear on the temp. control.  We know they're going to run hot 0.9V@2A. is temp control for overclocking? Or as nice-to-have for monitoring?

[FrankBuss]

You are using a linear regulator for 0.9V core voltage?

No, this would be too much energy waste. As written in my last message, I plan to use this module:
http://www.digikey.de/product-detail/de/ATS030A0X3-SRZ/555-1252-1-ND/2665523

That board says 5.6GH/s per board.  They are using different chips?

They are using the previous generation of the Avalon chips, the A3256. But should be not very different, only less fast hash calculation.

Would there be any advantage to keeping the avalon board separate from the PIC and the rest? I was thinking that having a 10-avalon board with just the serial and power. Then have the PIC and the rest on another board. This may make it possible to swap out the 10-avalon board easily and possibly give an upgrade path for the next-gen avalon.  Maybe make it more expensive but possibly offer reuse or repurpose even.

Good idea. And if I screw up the microcontroller part or if it doesn't work (I always forget some connections or something doesn't work as expected), it is easier to fix.

I'm not clear on the temp. control.  We know they're going to run hot 0.9V@2A. is temp control for overclocking? Or as nice-to-have for monitoring?

It is for monitoring. I think with a big heatsink, and if I use only 10 Avalons per board, no fan may be needed. But if the ambient temperature rises to high, it might need to shut down, if it gets too hot. If it gets really hot really fast, it can send you a SMS, informing you that your house is on fire

[scientist]

How many actual BTC do you typically get per GH?

[FrankBuss]

How many actual BTC do you typically get per GH?

This depends on how long you mine, see http://www.bitcoinx.com/profit/
Currently it is 0.0002 BTC per GH per day.

[scientist]

Wouldn't you get more $ if you mined Dogecoins instead? According to that link, your setup is going to be unprofitable for a very long time.

[FrankBuss]

You are right, it will need some months. As I wrote earlier, it is a just for fun project, and I can learn a bit from it.

For Dogecoins you'll need more powerful hardware, because of the required RAM for the algorithm. The Avalon chip can't mine it.

[GiskardReventlov]

That power supply part seems like the cheapest solution. Is it possible to find a part that will meet the specs for 15A? I guess that might add more complexity to the design too. My idea was to have two PS per module. But that may be more work than it's worth. However if you consider repurposing wouldn't a 15A supply be more useful than a 30A in other projects?

Another thought after looking at some of those PS parts and noticing that there's not that much to them. Maybe you could design one that exactly meets your requirement and save money.  But than they become a one-off and may not be reusable for much else.

[FrankBuss]

You are right, I'll use the power supply module, and don't build one myself. Someone wrote me a part which I could use, but the full circuit would cost EUR 13, not much less than the EUR 18 and it would not be as perfect as one of these modules.

There are lots of these modules, no need to design my own, see here.

[GiskardReventlov]

There are lots of these modules, no need to design my own

Definitely a lot simpler to buy the complete part.

As for the heat dissipation I see the options as:

1. heat sinks and fans to warm the house
2. heat sinks and fans to create waste heat
3. utilize the Seebeck effect to generate power to feed back into mining
4. utilize the Seebeck effect to generate power to charge batteries, cell phones, etc.
5. utilize the Peltier effect to cool the incoming are to the fans to reduce fan speed and noise
6. utilize the Peltier effect to cool your temperate zone saltwater or freshwater aquarium

For a nice example of the Seebeck effect in a consumer product look at the BioLite stove. I really love that thing and I hope Dave buys one and does a teardown.

[FrankBuss]

Meanwhile I have one of those power supply modules and tested it with a stripe of copper of an old PCB, because at 30 A the resistance of the copper gets interesting, and I have never done this before. It is 12 cm long and 1.4 cm wide, this is the test setup. The load resistance was just some 2 mm^2 flexible wire, cut until I got 46 milliohms, for 22 A at 1 V:



I measured 7.6 milliohm for the copper stripe. With 20 A this would result in a voltage drop of 0.152 V and a power dissipation of 3 W. The power supply module gets very hot, too, needed a fan.

The temperature of the copper stripe, shot with my new Fluke VT02:



Maybe I can use some wider traces and 70 um instead of 35 um copper. And the sense-pin of the power module helps a bit, but still a problem if one Avalon chip is up to 20 cm apart the other chip, which is important because of the low core voltage of 0.9 V. But this could be solved with soldering some silver wires on top of the traces. Just in case I'll remove the solder stop mask for it when creating the PCB. Maybe this is one reason that many other mining boards are using 4 layers?

Next week I'll get the PIC (it is available in SMD and DIP) to do some tests on a breadboard before I finish the Eagle schematic.

I don't think that the Seebeck effect would produce much more power than needed to light a LED, the efficiency is only 1-2%. Would be better to power some Stirling motor with the waste heat

BTW: the "SmartView" software for the VT02 is crap. First it didn't install on my PC and I had to search the temporary files to start the installation program manually. And when exporting an image, it doesn't export the scale on the right side properly (missing temperature numbers) and there are no useful features in the software at all. For example I would like to click with the mouse in the image and get the temperature, or adjust the image transparency overlay, as it is possible on the device. I wonder why the installation package needs 200 MB. But the device itself is nice, now they should hire some better programmers.

[GiskardReventlov]

I measured 7.6 milliohm for the copper stripe. With 20 A this would result in a voltage drop of 0.152 V and a power dissipation of 3 W. The power supply module gets very hot, too, needed a fan.

You are modeling the current/temperature profile of the PCB, right?

Maybe this is one reason that many other mining boards are using 4 layers?

I think you said 4-layer PCB is much more expensive. And you're saying that a 4-layer PCB allows multiple paths and therefore lower temperature and current per path?

Next week I'll get the PIC (it is available in SMD and DIP) to do some tests on a breadboard before I finish the Eagle schematic.

What hardware device do you use to program the PIC? I think I am going to go the atmel route to learn some mC programming.

I don't think that the Seebeck effect would produce much more power than needed to light a LED, the efficiency is only 1-2%. Would be better to power some Stirling motor with the waste heat

I thought an LED was super low power but the biolite (I haven't checked) must generate 5V 0.5A, I think, for charging.

BTW: the "SmartView" software for the VT02 is crap.

I guess you are on windows?  Some of the dev. env. on windows seem to allow (encourage?) bloaty code. But really not limited to windows. I switched to linux (lots of bloaty dog code here too) but way more choices. I would be surprised if there weren't some alternatives for the VT02 for linux, maybe some for windows too. Check github.

Reminds me of "What Intel giveth Microsoft taketh away."