DIY PCB with AVALON: "The Quarter Stick" - Needs Help!

[allten]

OK, I have a preliminary Schematic for engineering discussion.
https://www.dropbox.com/s/9c51uivncr5hhnp/DIY_AVALON_Rough_Draft1.sch.pdf

https://www.dropbox.com/s/ypqbam0uxwtcav9/BOM_Rough_Draft1.xls

Please take the engineering with a grain of salt. I would love to have some EE's look it over and
offer suggestions for improvement. I'm still having difficulty solving the math to yield values for
the following components:
C2?
C3?
R4?
C6?
c7?
L2?

Any help would be appreciated.

thanks,
     Allten

I'll join others here in saying that it's going to be much more cost effective to use a switching regulator with integrated MOSFETs. The NCP3170 looks appropriate.

As for C6, C7 and L2, these appear to form a low-pass LC filter. The corner frequency of such a filter is 1 / (2 * pi * sqrt(L * C)), and the filter is second-order. So for example, if L = 100 nH and C = 44 uF, the corner frequency is 75 kHz. This means the 642 kHz input ripple current is reduced by a factor of (642 / 75)^2 which is about 70. How low do you want the input ripple current to be?

Thanks for the feedback. You are right. that part is much better for the single avalon chip application.
I don't know how low the ripple current should be. I've looked through all the USB spec's and I couldn't find
anything that addressed this directly. The closes thing I found was related to the initial inrush current:
Maximum = 100mA/1uS

[BitHav]

Doe anybody have a broken board to lend? After a week I will return it fixed for free as a gesture of gratitude (except for the ASIC chips that still are not available).

Someone did post further up this thread that they had a broken board and wanted to help with figuring  out the details.

That was me BkkCoins

Hi Dieguito,

Where are you located?
Still have your broken Avalon Board?

[ecliptic]

Happy to join in on this, I've got some experience in PCB (& Schematic) Design and assembly.

It looks like the official avalon miners used a heatsink mounted on the opposite side of the board of the QFN chips and (persumably) a large number of thermal vias between the top and bottom layers?  I think i have heard 2W/chip thrown around as the TDP?

I'll have to catch on the thread tomorrow

Offhand i would say

* If the intent is at all to allow this to be soldered/assembled at home, don't use any QFN/BGA/etc packages.  The 48 QFN of the avalon chips themselves should be the most difficult task (but doable with the right soldering methods, possibly such as taping/holding the chip in place and drawing an iron with solder across the pads, the soldermask preventing them from bridging)
* Don't try to force the layout onto 2 layers.  4 Layers is the industry standard minimum for this sort of thing, and there is not significant price difference between 2/4 layer in volume.  Parasitics will be much better with power & ground plane, and routing everything on 2 layers could easily become a nightmare.  If the intent is to make it so you could produce this PCB with say a milling machine (and not have to order from a board house) frankly ... good luck.  Very difficult to mill out a board for all those QFN's and solder successfully if you can even get the traces the right size.  Good chance of pulling up copper traces in the thin areas, etc.  I'd say for certain the PCB needs to be made by a board house (group order will be orders of magnitude cheaper than individuals, and the board house may even be able to ship out to many different addresses if you work with them), and it may make sense to have a group order of the PCB assembly itself where you provide the assembly house with your PCB (or the gerbers/etc so they can order it), The Avalon chips, and they buy the rest of the parts, professionally assemble it and ship the units out.

Is avalon going to release a reference design?  That would be great, if so!

Oh, Yeah i'm assuming those QFN-48 chips have a big thermal pad on the bottom.  That means they're basically impossible to hand solder correctly (if you simply solder the pins you will get very bad heat transfer from the QFN package itself to the PCB, the internal PCB heatsinking on the presumably ground plane, and any heatsinking on the opposite side of the board. 

Theoretically it is possible to hand solder the thermal pad by putting paste on the top, and by having a large number of thermal vias going to the other side of the board, and then a big rectangular exposed copper surface that you can apply a soldering iron to to transfer heat up, reflow the paste, and have the chip fit into position.  Not nearly as ideal as mass produced pick & place reflow, but possible.

[dieguito]

Doe anybody have a broken board to lend? After a week I will return it fixed for free as a gesture of gratitude (except for the ASIC chips that still are not available).

Someone did post further up this thread that they had a broken board and wanted to help with figuring  out the details.

That was me BkkCoins

Hi Dieguito,

Where are you located?
Still have your broken Avalon Board?

Hi BitHav I'm located in Buenos Aires and I think BkkCoins is in Asia isn't he?

[samurai1200]

@BitHav

Where you are getting 5V that is feeding the 3.3V LDO? USB?

If 3.3V is using any more than 250mA, I would consider using a buck regulator supplied by 12V as well. No need to create more heat with an LDO (however little in comparison).

[ecliptic]

I see BURNIN's design is supposed to be open source, that is something worth looking into when it is released.

[BitHav]

@BitHav

Where you are getting 5V that is feeding the 3.3V LDO? USB?

If 3.3V is using any more than 250mA, I would consider using a buck regulator supplied by 12V as well. No need to create more heat with an LDO (however little in comparison).

PIN 20 of the board connector is 5 V input.
The complexity and cost of a Buck converter is too overwhelming for just a few miliwatts. Supposing 250 mA, times the voltage drop (5-3.3= 1.7 V) is a PD of .25*1.7 = 0.42W and I thing it is much less than that.

[samurai1200]

@BitHav

Where you are getting 5V that is feeding the 3.3V LDO? USB?

If 3.3V is using any more than 250mA, I would consider using a buck regulator supplied by 12V as well. No need to create more heat with an LDO (however little in comparison).

PIN 20 of the board connector is 5 V input.
The complexity and cost of a Buck converter is too overwhelming for just a few miliwatts. Supposing 250 mA, times the voltage drop (5-3.3= 1.7 V) is a PD of .25*1.7 = 0.42W and I thing it is much less than that.

Ohh i see what youre saying... "the board" being the backplane of the avalon?

[BitHav]

@BitHav

Where you are getting 5V that is feeding the 3.3V LDO? USB?

If 3.3V is using any more than 250mA, I would consider using a buck regulator supplied by 12V as well. No need to create more heat with an LDO (however little in comparison).

PIN 20 of the board connector is 5 V input.
The complexity and cost of a Buck converter is too overwhelming for just a few miliwatts. Supposing 250 mA, times the voltage drop (5-3.3= 1.7 V) is a PD of .25*1.7 = 0.42W and I thing it is much less than that.

Ohh i see what youre saying... "the board" being the backplane of the avalon?

Yeah, actually pin 20 in both boards...

[samurai1200]

 Right,  yeah. I  was  mostly just curious about the origin of 5v.

[aeronautical]

I mentioned in that a standard schematic and pcb should be used in some of my posts for development work and said that Eagle Cad was the best.  I noticed that some have used Eagle Cad in posts. I found later that all the designs were done using Altuim Designer. http://www.aeronautical.com.au/resources/Dual%20Spartan6%204%20BTC.png  Here a image link to one of the PCB’s that we had up on Altium Designer

[BkkCoins]

PIN 20 of the board connector is 5 V input.
The complexity and cost of a Buck converter is too overwhelming for just a few miliwatts. Supposing 250 mA, times the voltage drop (5-3.3= 1.7 V) is a PD of .25*1.7 = 0.42W and I thing it is much less than that.

I just thought I'd throw a comment in here. In my case since I want to use PCIe power I won't have an easy 5V supply, without adding another cable. But also, the price difference between a buck and LDO isn't that significant for lower power use. With 12V input it's a better choice here.

I was guessimating that each ASIC may use 25-50mA each. Probably it's just for I/O driving but it could also be used for something else (the PLL?, clock osc is 3.3V). Not being sure I went with the higher value. On a 10-16 chip board that would be 500-800 mA.

At 12V we have (12-3.3) * 0.8 = 7W dissipated if using an linear reg. Now the one specified in Avalon is rated at 600mA AP7215 so we can guess he chose a suitable part for 10 chips. But that reg costs around $1.50 qty 10 @ Mouser.

Using an MCP16321 we can have 12V input, rated 89% efficiency, gives 0.26W loss. An MCP16321 costs $1.55 qty 10, and you need a 4.7uH inductor that costs $0.27.

So while at 5V it's feasible at 12V it's a bad idea even if you can get rid of the 7W heat.

[SebastianJu]

It seems nobody answered my question regarding a board with higher amount of chips so i will ask again.

Whats the best performance in regards to how many chips are on a pcb? I believ avalon is putting 16 on a pcb, burnin thinks about 20 chips. But i wonder wouldnt it be more efficient to make a big pcb with maybe 100 or 200 chips on it? Only one step to implement many chips, probably one power supply and so on.

I mean i see how many chips some people order. And i see the prices for "one board". I dont believe that a chip with 1000 chips would be cheaper at all because of the size that would drive the production cost into heights. But where is the optimum here? Burnin says 10 chips for 80€ and 20 for 100€. So its cheaper to build in more chips. Wheres the possible optimum?

I really think the one that develops this "cheapest" board will make a fortune. And the one that creates the biggest board will be friend of those that bought a batch on its own and dont want to have staying around 1000 miners with 10 chips each. The more units the more costs, the more needed connections and so on.

[Bicknellski]

It seems nobody answered my question regarding a board with higher amount of chips so i will ask again.

Whats the best performance in regards to how many chips are on a pcb? I believ avalon is putting 16 on a pcb, burnin thinks about 20 chips. But i wonder wouldnt it be more efficient to make a big pcb with maybe 100 or 200 chips on it? Only one step to implement many chips, probably one power supply and so on.

I mean i see how many chips some people order. And i see the prices for "one board". I dont believe that a chip with 1000 chips would be cheaper at all because of the size that would drive the production cost into heights. But where is the optimum here? Burnin says 10 chips for 80€ and 20 for 100€. So its cheaper to build in more chips. Wheres the possible optimum?

I really think the one that develops this "cheapest" board will make a fortune. And the one that creates the biggest board will be friend of those that bought a batch on its own and dont want to have staying around 1000 miners with 10 chips each. The more units the more costs, the more needed connections and so on.

With such an order you might be better off asking one of the DIY PCB designers to work with you specifically for a larger board that would be more efficient in terms of components and power compared to smaller offerings if it is even possible. Can't see why it wouldn't. Or find someone to develop a design for you. There are sites where you can hire designers for this https://www.odesk.com/.

[BkkCoins]

It seems nobody answered my question regarding a board with higher amount of chips so i will ask again.

Whats the best performance in regards to how many chips are on a pcb? I believ avalon is putting 16 on a pcb, burnin thinks about 20 chips. But i wonder wouldnt it be more efficient to make a big pcb with maybe 100 or 200 chips on it? Only one step to implement many chips, probably one power supply and so on.

I mean i see how many chips some people order. And i see the prices for "one board". I dont believe that a chip with 1000 chips would be cheaper at all because of the size that would drive the production cost into heights. But where is the optimum here? Burnin says 10 chips for 80€ and 20 for 100€. So its cheaper to build in more chips. Wheres the possible optimum?

I really think the one that develops this "cheapest" board will make a fortune. And the one that creates the biggest board will be friend of those that bought a batch on its own and dont want to have staying around 1000 miners with 10 chips each. The more units the more costs, the more needed connections and so on.

With such an order you might be better off asking one of the DIY PCB designers to work with you specifically for a larger board that would be more efficient in terms of components and power compared to smaller offerings if it is even possible. Can't see why it wouldn't. Or find someone to develop a design for you. There are sites where you can hire designers for this https://www.odesk.com/.

With my design it will be possible to panellize multiple 10x10cm boards onto bigger ones. But it doesn't get much more economical to do that and you lose some modularity. A 20cm x 20cm board is quite reasonable and that would hold 4 modules with no extra design effort - just plunk the gerber files down 4x. And you end up with 64 chips / board. Or 20x30 would give you 80 chips, which is nice as it's about the same size as a sheet of paper, so still easy to work with.

I should note if you panellize then you can leave off the usb connector except for one per big board. You would make an octopus PCIe power splitter joining one PSU lead to one big board, with 4 or 6 plugs. And a small 3 pin ribbon jumper to hop around the modules for I2C data.

[Bicknellski]

I'll take 7 of your 20 x 30 Boards BKK lol.

[BitHav]

PIN 20 of the board connector is 5 V input.
The complexity and cost of a Buck converter is too overwhelming for just a few miliwatts. Supposing 250 mA, times the voltage drop (5-3.3= 1.7 V) is a PD of .25*1.7 = 0.42W and I thing it is much less than that.

I just thought I'd throw a comment in here. In my case since I want to use PCIe power I won't have an easy 5V supply, without adding another cable. But also, the price difference between a buck and LDO isn't that significant for lower power use. With 12V input it's a better choice here.

I was guessimating that each ASIC may use 25-50mA each. Probably it's just for I/O driving but it could also be used for something else (the PLL?, clock osc is 3.3V). Not being sure I went with the higher value. On a 10-16 chip board that would be 500-800 mA.

At 12V we have (12-3.3) * 0.8 = 7W dissipated if using an linear reg. Now the one specified in Avalon is rated at 600mA AP7215 so we can guess he chose a suitable part for 10 chips. But that reg costs around $1.50 qty 10 @ Mouser.

Using an MCP16321 we can have 12V input, rated 89% efficiency, gives 0.26W loss. An MCP16321 costs $1.55 qty 10, and you need a 4.7uH inductor that costs $0.27.

So while at 5V it's feasible at 12V it's a bad idea even if you can get rid of the 7W heat.

Of course, if you don't have 5V available, then the more efficient solution is what you are mentioning. The topology to get a solution is always an engineering decision, and you have good arguments.

Here is a link to a possible solution that per TI webench  cost only $1.32 (without PCB) in a pcb area of only 127 mm2. See details here:

http://floridaposts.com/PDF/3.3V-1A-BuckConverter.pdf

BitHav

[BitHav]

It seems nobody answered my question regarding a board with higher amount of chips so i will ask again.

Whats the best performance in regards to how many chips are on a pcb? I believ avalon is putting 16 on a pcb, burnin thinks about 20 chips. But i wonder wouldnt it be more efficient to make a big pcb with maybe 100 or 200 chips on it? Only one step to implement many chips, probably one power supply and so on.

I mean i see how many chips some people order. And i see the prices for "one board". I dont believe that a chip with 1000 chips would be cheaper at all because of the size that would drive the production cost into heights. But where is the optimum here? Burnin says 10 chips for 80€ and 20 for 100€. So its cheaper to build in more chips. Wheres the possible optimum?

I really think the one that develops this "cheapest" board will make a fortune. And the one that creates the biggest board will be friend of those that bought a batch on its own and dont want to have staying around 1000 miners with 10 chips each. The more units the more costs, the more needed connections and so on.

With such an order you might be better off asking one of the DIY PCB designers to work with you specifically for a larger board that would be more efficient in terms of components and power compared to smaller offerings if it is even possible. Can't see why it wouldn't. Or find someone to develop a design for you. There are sites where you can hire designers for this https://www.odesk.com/.

With my design it will be possible to panellize multiple 10x10cm boards onto bigger ones. But it doesn't get much more economical to do that and you lose some modularity. A 20cm x 20cm board is quite reasonable and that would hold 4 modules with no extra design effort - just plunk the gerber files down 4x. And you end up with 64 chips / board. Or 20x30 would give you 80 chips, which is nice as it's about the same size as a sheet of paper, so still easy to work with.

I should note if you panellize then you can leave off the usb connector except for one per big board. You would make an octopus PCIe power splitter joining one PSU lead to one big board, with 4 or 6 plugs. And a small 3 pin ribbon jumper to hop around the modules for I2C data.

You guys are dead on the money.

I am working my own design with 16X16 Avalon chips in a single board, using a PCI-ex4 interface, and the standard PCI-e 12V suplemental power as additional power source. If all the footprint is there (and much more powerfull) in an ATX motherboard, why bother building additional boards and interfaces as in the Avalon Design?

With 256 chips per board you can hash conservatively 64GHs/s per board. You can stack up to 5 of these boards together for over 320 Ghs/s per unit.

Worth to try...

[BkkCoins]

With 256 chips per board you can hash conservatively 64GHs/s per board. You can stack up to 5 of these boards together for over 320 Ghs/s per unit.

Worth to try...

And also each board could double as a pro grade waffle iron...

[||bit]

I wonder what the power cost would be to use a fuel cell instead of line power. I know it would be hard to beat power company rates, then there's the cost of a fuel cell. Advantage of a fuel cell would be that it would be easier to regulate the voltage and it should be nearly flat (cleaner) since there would be no ripple. Perhaps, this cleaner power would allow for lower amplitude clocking, and so less power consumed. Less heat should make room for a hair more up-clocking. I also wonder if cleaner power having, b/c of less noise, would also allow for any higher clocking. Not sure.