No, I was meaning the defective microcontrollers that are delaying the whole thing while marto tests them and swaps by hand the ones that doesn't work.

I can do it myself if they are not bga and perhaps if some of us can do the same marto will finish the others faster = everyone wins.

Someone that has the boards, can we get a close up of the microcontroller?

If it is something that I can replace I wouldn't mind doing it myself as sidehack.

Interesting spreadsheet. Can you try 400mhz at 860mv?

It should burn about 294W by my numbers. It should also hash at the same speed and the efficiency would be shit but if it burns way more or way less than that it might tell us where to look.

Are you using it for a single board?

It is a great psu and I have 8 of them waiting but a single undeclocked board might be a too small load, see here (it is for the 2500W one but they are similar):

http://www.rcgroups.com/forums/showpost.php?s=ff664b0f8735ce2a9c40796f40b57c49&p=18117345&postcount=611


It would be great if you could measure volts and amps at the 12V rail or if you could add more load to the psu if you are using one board.

If that is not the case it is strange, crappy voltage regulators should get way more inefficient as the load increases exagerating the advantage of the underclocking so it can't be that. Having less capacitors might increase the ripple, overshoot and undershoot but it shouldn't wreak havok with the efficiency. The only way I can picture the dc-dc converters drawing a lot of extra power due to having less capacitors and it being more noticeable with more underclocking is if marto increased that dc-dc converter frequency a lot to compensate for the loss of capacitance and that forced him to increase the gate drive of the mosfets to insane levels wasting a ton of power driving them. I have done that wasting up to 20W per mosfet in a high voltage generator but it required a discrete gate drive circuit, I have never seen anything like that in computer hardware and it would be insane if you cared about efficiency or even cost as caps are cheaper than building discrete gate drives.

So There isn't any machine that can desolder them.

Then why not making it, perhaps making it fully automated is not an option but what about machining a copper tip with the exact shape for the mcu. A small hole for a thermocouple and series of channels to connect a venturi to suck the solder away. Even doing it by hand that way you could replace hundreds/day.

And to test them you could also make some kind of contraption to speed up things.

In fact being abble to work with metal you could also make frames for the 4 heatsinks that doesn't even need screws so adding a bit of thermal paste is all that is needed to change the boards for testing. If you need it I can make a drawing of what I mean.

I am not disputing your expertise but maybe some fresh ideas would help speeding up things.

If so many mcus are defective wouldn't it make sense to have them all removed and swapped for new ones by a machine and leave the testing for a later time? (I don't know if there is a desoldering machine)

Even 37 boards/day is painfully slow with the total number of boards in mind and it must be quite tiring to do it all by hand time and time again.

Does anyone know how many boards you can hook up to a single tp-link?

I will have to set up 25 boards (mine an a friend's). We already have more than enough 8 port usb hubs that work realiably in series.

I have one tp-link and I already flashed the image while waiting, the interface does not give any indication about the number of boards and there are just one set of fields for the configuration so will it just try to set up all the boards it finds with the same parameters?

I have searched the forums but there is no conclusive number, I have seen some people claiming up to 10.

Artic silver epoxy could be a way to fix it although it is too permanent for my liking.

Martin has metalworking machinery so I think that a clip like the ones I made for my watercooling could fit most aircoolers too. It is just an H with two 1cm thick aluminium bars and a threaded 4mm steel bar in the middle, it is strong enough and being so thin it should fit in most coolers.
I already made mine with 75mm centers so I hope that there is enough leeway to fit my clips perhaps by using thinner bolts so they can be angled a bit more.

WTF! Removing caps like that can easily leave the pieces of the leads attached to the pcb and cause a massive short.

+1 , I have the ultra and it dropped about 8ºC on my cpu and 15ºC on my graphics card.

However as it has been said, it is gallium based and gallium destroys aluminium :

https://www.youtube.com/watch?v=JHHI2Lk79cY

Also it is expensive and the improvement in cooling depends a lot on the watts/cm^2 of the die, I don't know if it is really worth it for the minion so I will be using artic silver.

BTW my watercooling tests are going fine, a relativelly complex block that could be used as a phase change evaporator gave me 28ºC with 100W (a p4 northwood) with 22ºC of ambient temps. However I can't really mass make those on time or budget so I tried a really simple high flow design and I got 38ºC load temps wich even if it is not stellar performance it is good enough so I am making 100 of them (for me and a friend). I also got a monster 155W 10000L/h pump.

It is a missconception that watercooling would be expensive, it only is when you buy overpriced computer cooling components. I am making the waterblocks myself with a drillpress and copper bar, it is a bitch to drill and if the shape is complex it can take quite a bit of time and skill but fortunatelly even a straight drill performs relativelly well. You can also use aluminium, it would cost almost nothing, easy to drill and it would still outperform most aircoolers.

There is also an economy of scale, I am going to use a single 10000L/h pump (costs about the same as pumps meant for computer cooling) and a big car radiator. Also instead of a ton of fans I will place the radiator over an extractor fan and that would also remove the problem of heat buildup in the room.

In the end it will be cheaper than air heatsinks and perform beter.

Hi, could we get precise information on board dimensions and component placements? (at least the asics).

I am going to watercool them and if I know the measures with precission I would have a lot in fittings and use soldered copper pipe that I already have and just 2 fittings per board.

Also it would help for making cases/stacking them/planning the space needed.


Thanks.

I am just going by my 360mm heater core radiator performance, I have dumped 1KW on it by watercooling a condensor before without problems. It is a matter of having the right water and airflow. I had 4 fans in pushpull so I was just using 240mm of the radiator and a 1000L/h pump and was just about 15ºC over ambient....

Those radiators are really nice, even only one should perform quite well with 4 minions but I wouldn't use that pump. If you want to save more or anyone else is building a watercooling, a car heater radiator performs just as well as the most expensive computer radiators.

A koi pond pump would beat the crap out of it for about the same price. Just one of those pumps could service 30 minions in a properly designed setup.

I am developing a dual purpose phase change evaporator / waterblock and if you are interested I could hook you up with a prototype.
The current one is underperforming in the flow ressistance tests with water but I have narrowed down the problem and I am working on a solution.

I hope to reach my performance targets soon but meeting the needs of phasechange, watercooling and having a cheap and easy to assemble design is hard.

I think that watercooling is quite practical and a good investment due to the reuse factor, I have been using the same pump and radiator since year 2000 and even my original aluminium waterblock would have suficed with reduced performance but stil better than air.
In a crowded datacenter it would make even more sense as you could reduce the air conditioning bill by a lot if you place the radiator outside and you could use car or truck radiators with a nice and big pump (or several in parallel to increase the reliability).  It can also be cheap, as I have said the pumps and radiators can be shared and cheaper that those sold by computer companies and if you diy the waterblocks it is not expensive at all.

The compensation is awfully inadequate unless we take the just give me the chips route and then it means trusting someone else to assemble them, extra delays, laying out extra money and so on.

But I have one idea that might work for us and for black arrow. Add free minions to the compensation package, I am sure that you can make them for cheap, we could try selling them or sending them to technobit.

I understand that any sort of engineering project can see delays and extra expenses so I don't doubt your word that you just can't give us more X3s but some chips wouldn't be such a burden as a fully assembled miner and it would make a difference for us and might somewhat save your public image.

Of course I can't be sure but software control of vcore is quite standar in modern computers and problems appart it is clear that prosperos are beautifuly designed so it wouldn't surprise me if they had it. At the very least they should have it but not exposed in the interface if just because it is a convenient way for them to do a late change without changing resistors. Anyway if they don't have it or if the range is insuficient I would also like knowing wich pwm controller they use ;-).

The psu might limit the oc but I am meaning to go sub 0 and I hope that means a significant reduction in leakage and I hope that it might be enough to offset the increase due to higher frequency (if those things love vcore then we are screwed as the relationship with power draw is quadratic at the least).
Also it would be hard finding more powerful power suplies, it would be expensive and if you have to pay electricity it would be too expensive if it is going to be over 4500W per X3.

I hold hopes of getting answers because it is David's personal miner. And I don't see why wouldn't he want to test it and share the results with us.

david105396 as you will get your miner soon I am very interested in some information.

The relationship between temperature and overclocking.
The relationship between vcore (not me lol) and overclocking.
The relationship between temperature and power consumption.

If there is enough overhead in the stock heatsink you could test it by controlling the fan manually and setting it to the maximum. If there isn't it will also be interesting.

My guesses are that overclocking will be heavily dependent on temperature a midly on voltage hitting thermal limits with slight increases on vcc while on stock cooling, I also think that it is quite likelly that there is a significant decrease in power consumtion with decreased temperatures. For those things to be really noticeable we need at least 20ºC of difference.

That information would really help us extracting the best overclocking and or getting a better hash/w. I am building a custom cooling and it would help me quite a bit.
Also, what is the maximum overclocking and voltage that the interface/hardware lets you set? (even if it will never reach it with aircooling it might with water or phase change and what could be reached could determine what kind of cooling is worth).

If the problem is noise in the 3.3V line it is easy to test it.

Cut the 3.3V traces going to the pcie conectors in a backplane or if it cant be done due to the layout/multilayer pcb just remove the 3.3V pads in the hashboard and then feed it with clean 3.3V, if you want really clean 3.3V use a 3.3V linear regulator feed with at least 5V and with big filter capacitors at the input and output.

Not the most efficient and likelly not an aceptable solution but the power will be as clean as it can be and you will know for sure if it is worth trying to make a pcb that has lower noise on the 3.3V line or not.

That can be cobled together in less than an hour if you already have the linear regulators. I like this one: http://www.ti.com.cn/cn/lit/ds/symlink/lm1085.pdf .

Although perhaps that is exactly how you came to the idea that the problem is in the 3.3V line.

Surely most of those 96 pins are power and the backplane doesn't have any problem with power does it?
So why not cutting the spi traces and running wires only for the spi?
It might not be elegant but depending on the pcb it might be workable (ie, desoldering something and putting an spi only connector glued to the board) or at least it works to test it.

I am fine with leaving the lcd out as long as it doesn't affect the minning performance, reliability or overclocking. Perhaps you should consider excluding the lcds and giving the psu for the compensation units free of charge.

Thanks.