Server Power Supply Interface Board - for standalone miners and GPU rigs

[Mk2vr6]

Preemptive disclaimer that I don't work with GPUs a lot.

But I would imagine the 300W and 280W from your cards, part of that is coming from the risers. If you don't plug the risers into a heavy external supply but run them off your ATX, I'd allot at least 6A of 12V per riser. If the 300W/280W is straight external, they should all run off the DPS-1520 without issue. If that includes the socket power (from risers in this case), you could power the risers off the DPS-1520 and still only be at 80% capacity.

If, then, you ran your 6950 and motherboard off the ATX supply, I'd recommend probably at least a 500W unit. Maybe more depending on what processor/motherboard you're running, and what overhead you want to allow for.

What you could do, and an idea I've toyed around with but haven't tested yet, is actually get two server supplies in parallel load-balanced (I don't know enough about that particular model to know if they current-share or not, but I think most do) and get a picoPSU running off the 12V bus for your motherboard. Those things aren't terribly expensive, and run around 96% efficient. If you isolate your riser power from the motherboard power (either by using the recommended USB-style risers or cutting the 12V lines to the ribbon) you shouldn't overload anything on the picoPSU by trying to pull GPU current through it, and anything else requiring 12V (like GPU, or processor VRMs) would pull straight from your ~3KW 12V source.

Right, I've somewhat conflicting info from elsewhere now.

Simplified - I have DPS-1520 wired to 4x GPU's. Can my Desktop PSU power the MOBO and the powered risers too (plus an extra card if I have the headroom) ?

[sidehack]

Depending on the strength of your desktop supply, it should. Powered risers you should allow at least 75W of 12V per riser, and at least 300W of 12V for your other GPU. Depending on your board/CPU, you'll want 100-200W for that. Looking at getting at least 800W desktop supply to run all that.

If anyone, ever, anyone at all, has any actual experience building GPU rigs with riser cards, feel free to chime in. The most I've done is 3 GPUs direct on a board, with an ATX giving 35A of 12V and a 500W server supply doing the rest. The only advice I can give regarding high-end crap is somewhere between theory and conjecture. Never used risers, never used an ATX supply over 500W, never used more than 3 GPUs, never used a DPS-1520.

[Mk2vr6]

Depending on the strength of your desktop supply, it should. Powered risers you should allow at least 75W of 12V per riser, and at least 300W of 12V for your other GPU. Depending on your board/CPU, you'll want 100-200W for that. Looking at getting at least 800W desktop supply to run all that.

If anyone, ever, anyone at all, has any actual experience building GPU rigs with riser cards, feel free to chime in. The most I've done is 3 GPUs direct on a board, with an ATX giving 35A of 12V and a 500W server supply doing the rest. The only advice I can give regarding high-end crap is somewhere between theory and conjecture. Never used risers, never used an ATX supply over 500W, never used more than 3 GPUs, never used a DPS-1520.

Still a damn site better than knowing nothing - Thanks buddy

Two more questions while I've got your attention, if you don't mind?

-Would I be able to wire up MOLEX connectors aswell as the PCIE to the DPS-1520 to power the four cards/risers. And then use say a 500-600w PSU for the other card/riser/mobo?

-Is it wise to run two lengths of both live/earth from PSU to each 6 pin connector per card, or can it be wired so that the second 6 pin is jumped off the first?

Like this -

[sidehack]

Would you be able to? I think that questions your capabilities as a hardware assembler and I can't answer that for you.
But anyways... If the risers only take in 12V from the molex, and the cards you mentioned pull the power you mentioned combined through the 6/8-pin and risers, then yes the DPS-1520 should be able to power your 4 GPUs and their attached risers. Your ATX would then only need to source power for the other GPU/riser (~300W?) and motherboard (~200W?) so a 500/600 ought be sufficient. Probably?

I'd rather run wires for each connector instead of sharing. It can be wired with two connectors sharing the same wire, but its reliability depends on the length and gauge of the home run because you'll be pulling the full current of both connectors through it. The more current through a wire, the more power dissipation in the wire. That means the wire warms up and, if it's way under spec, could melt or burn something. Power dissipation in the wire is also a result of voltage drop in the wire, which means less voltage getting to the load, which means a higher current required to power the load fully, which further increases the strain on your wires. If it's calculated out properly you shouldn't have any problems, but in general the more copper the better (at least as far as is practical with weight/price, like for most of what you're doing 000AWG won't do you any better than, say, 14AWG).

[Mk2vr6]

Would you be able to? I think that questions your capabilities as a hardware assembler and I can't answer that for you.
But anyways... If the risers only take in 12V from the molex, and the cards you mentioned pull the power you mentioned combined through the 6/8-pin and risers, then yes the DPS-1520 should be able to power your 4 GPUs and their attached risers. Your ATX would then only need to source power for the other GPU/riser (~300W?) and motherboard (~200W?) so a 500/600 ought be sufficient. Probably?

I'd rather run wires for each connector instead of sharing. It can be wired with two connectors sharing the same wire, but its reliability depends on the length and gauge of the home run because you'll be pulling the full current of both connectors through it. The more current through a wire, the more power dissipation in the wire. That means the wire warms up and, if it's way under spec, could melt or burn something. Power dissipation in the wire is also a result of voltage drop in the wire, which means less voltage getting to the load, which means a higher current required to power the load fully, which further increases the strain on your wires. If it's calculated out properly you shouldn't have any problems, but in general the more copper the better (at least as far as is practical with weight/price, like for most of what you're doing 000AWG won't do you any better than, say, 14AWG).

Yes I am more than capable of putting it all together, but thanks for picking apart my question when the intention was plain to see :p

No, seriously, you've answered everything I need to know. I will get some Molex connectors tomorrow and put it all together.

And for the record I'm using 16awg wire, direct from China! Haha. Although it looks suspiciously thin for my liking, and I'd guess its actually 18awg. Either way I'll wire up the connectors separately to be on the safe side.

Thanks again for your time.

Tony.

[moca]

For anyone who want to save some time, i have fully modded PSU for sale.
https://i.imgur.com/0oAWMzI.jpg

IBM DPS-1300BB 1300w with 4x6 pin + 4x(6+2)pin plus the essential synchron-start-cable(24 pin ATX-->24 pin ATX+ 2 pin + 2 pin), with what the power on and off will be synchronized with the main ATX PSU.

Easy expendable to 2600w, just put another 1300w to the spare 2 pin from the start cable.

Here is the link where you can see more photos and details

http://www.ebay.de/itm/321349206895
(ye i know it's german, i live in germany)

If you are interested, pm me.

[TheBackRoads]

Question about the board, is the dial suppose to control the PSU fan speed on the 0.4 version?  From what I can gather, the dial works fine, and a reading out to a pin on the psu connector, but the fan speed never changes.  Any ideas?  Otherwise works great!  Love it!

[sidehack]

Every supply I tested, the voltage range output by the knob worked. However, HOWEVER... someone at one of Dell's various Chinese outsource design houses decided it would be a good idea to make some with a different reference range for fan speed. I guess it doesn't matter a whole lot if you have a closed-loop regulation system, but these boards don't do that.
If you have the skills to do so, shorting around the 393 resistor on the bottom side of the board should give you the range to control the fan speed on those supplies. This problem wasn't known until recently, so a true fix won't be applied until the V0.6 boards.
The other option is to wire a potentiometer between 3.3V and GND pins on the header, tap on the FAN pin, and set the board to use external fan control.

[TheBackRoads]

If you have the skills to do so, shorting around the 393 resistor on the bottom side of the board should give you the range to control the fan speed on those supplies. This problem wasn't known until recently, so a true fix won't be applied until the V0.6 boards.
The other option is to wire a potentiometer between 3.3V and GND pins on the header, tap on the FAN pin, and set the board to use external fan control.

THANK YOU!!! That worked perfect, I will for sure be buying these from now on.  You rock!

[sunbreak]

Preemptive disclaimer that I don't work with GPUs a lot.

But I would imagine the 300W and 280W from your cards, part of that is coming from the risers. If you don't plug the risers into a heavy external supply but run them off your ATX, I'd allot at least 6A of 12V per riser. If the 300W/280W is straight external, they should all run off the DPS-1520 without issue. If that includes the socket power (from risers in this case), you could power the risers off the DPS-1520 and still only be at 80% capacity.

If, then, you ran your 6950 and motherboard off the ATX supply, I'd recommend probably at least a 500W unit. Maybe more depending on what processor/motherboard you're running, and what overhead you want to allow for.

What you could do, and an idea I've toyed around with but haven't tested yet, is actually get two server supplies in parallel load-balanced (I don't know enough about that particular model to know if they current-share or not, but I think most do) and get a picoPSU running off the 12V bus for your motherboard. Those things aren't terribly expensive, and run around 96% efficient. If you isolate your riser power from the motherboard power (either by using the recommended USB-style risers or cutting the 12V lines to the ribbon) you shouldn't overload anything on the picoPSU by trying to pull GPU current through it, and anything else requiring 12V (like GPU, or processor VRMs) would pull straight from your ~3KW 12V source.

You mean like this; This is 3 systems all connected to the same DPS-2000BB power supply. Each system has 6 x PowerColor 6950's (some unlocked) for a total of 18 GPU's. This was an undervolted setup with the memory downclocked to 150mhz through a bios mod. So overall power consumption per system was about 750 watts. The IBM PSU if properly cooled will do 200 amps before OCP trips. So you can actually get upwards of 2300+ watts reliably.



A previous client and I had these picopsu like boards made in china in bulk. They are a special design, unlike a normal picopsu that regulates the 12V line and limits it's current. This design passes the 12 volt straight through without regulation and provides a full 10 amps on 5v and another 10 amps on 3.3v. The connector and wire gauges are sized so that the current from the 3.3v and 5v do no impact the overall current available on the 2 x 12 volt atx pins. Also the CPU connector is straight pass thru. The connector was spec'd so that I could be put on powered from a standard ATX PSU from the 8 pin CPU header (only 4 pin CPU pass thru). Since most PSU's have two 8 pins (one dual 4 pin) you could power two individual systems and then yet one more from the ATX PSU's monster 24 pin cable. However it's very difficult to have enough power to power 3 systems on a ATX PSU as most are limited to 1200-1300 watts.



This is a break out board. So the PSU powers each system with a single pair of #6 copper wire. Then this break out board gives you all the pig tails in addition to the 8 pin CPU style connector to power the motherboard.



A closer look at the PSU adapter break out I made.

[sidehack]

Nifty. I don't have any picoPSU running around but a future project I have in mind will incorporate one running off probably a DPS-800 supply.

We're working on DPS-2000BB breakout boards now, I hope to have a final schematic tomorrow and start on the PCB design. Should be interesting. Gonna build about a 3KW adjustable dummy load to stress-test them; the board should be rated for at least 200A without issue. It'll also incorporate a PWM fan speed controller with internal and external adjustment.

We're also working on a plug-and-go board/cable kit that'll allow you to load-balance and simultaneously control 3 boards at a time. It'll be good for the current D750 boards, and the DPS-800 and DPS-2000 boards we'll be releasing soon. The ability to turn all supplies on and off as a unit, or disable specific supplies in case of problems. The ability to daisy-chain controlboards for more than 3 supplies at a time. Status LEDs for PSU standby power, PSU-on and 12V present on the bus. And unified fan speed control. Should be pretty nifty, and solve a lot of questions people have been asking about parallel/load-balance. I'm all for people learning how to do things themselves, but some folks just can't, or won't, and then if they got an extra ten bucks or whatever maybe they'll just use the it's-already-done-for-you kit.

[sunbreak]

I guess the point I was trying to make was that the PicoPSU won't cut it. There is a lot of load on the 12 volt side of a pico psu. They just can't keep up with demands from the PCI-e bus let alone current from the CPU 4 pin header.

I'll have to dig up the details of who we used to manufacture. But it has been a reliable design and supports full ATX spec currents from a 8 pin cpu style header including dedicated 4 pin CPU header.

I think I still have like 20 or 30 of those 12v->atx+4pin-cpu adapters in the pic kicking around.

[sidehack]

Gotcha. They do also have different power outputs and stuff from different models, but I haven't actually messed with them. It'd probably not be too difficult to disable/hack the 12V line if necessary, but then again "difficult" is an intensely subjective qualifier.

[Mk2vr6]

I guess the point I was trying to make was that the PicoPSU won't cut it. There is a lot of load on the 12 volt side of a pico psu. They just can't keep up with demands from the PCI-e bus let alone current from the CPU 4 pin header.

I'll have to dig up the details of who we used to manufacture. But it has been a reliable design and supports full ATX spec currents from a 8 pin cpu style header including dedicated 4 pin CPU header.

I think I still have like 20 or 30 of those 12v->atx+4pin-cpu adapters in the pic kicking around.

Very Impressive setup.

Can I ask what each 6950 hashes at because I will be using a few, and Leccy IS NOT cheap in the UK.

Regards.

[jlsminingcorp]

Every supply I tested, the voltage range output by the knob worked. However, HOWEVER... someone at one of Dell's various Chinese outsource design houses decided it would be a good idea to make some with a different reference range for fan speed. I guess it doesn't matter a whole lot if you have a closed-loop regulation system, but these boards don't do that.
If you have the skills to do so, shorting around the 393 resistor on the bottom side of the board should give you the range to control the fan speed on those supplies. This problem wasn't known until recently, so a true fix won't be applied until the V0.6 boards.
The other option is to wire a potentiometer between 3.3V and GND pins on the header, tap on the FAN pin, and set the board to use external fan control.

Hi Sidehack, these are great boards and work fantastically, thanks. However, I think I have a related problem to this. I get some control of the fan speed from the board, but this only varies from hurricane (unbearably loud) to stormy (annoyingly load) and it would be great to get the speed down further, as I'm sure that the PSU doesn't need that level of air movement to cool it at the loads that I'm using. Apologies for the stupid question (I'm an electronics noob), but what kind of pot (what resistance?, linear single gang or something else?) do I need to wire up to use external fan control as you describe above?

[precrime3]

Hows this different than what gekko science is offering? Also what would be the best option to power 3 antminers off 1 psu and run them at stock clock.

[joeventura]

Hows this different than what gekko science is offering? Also what would be the best option to power 3 antminers off 1 psu and run them at stock clock.

This IS gekko science.

[sidehack]

External fan, the actual pot resistance doesn't matter a whole lot. I'd probably use a 10K. What really matters is a wiper voltage range that includes 0-1.8 or so.

Yes, GekkoScience is my business. Hence why all the links are to stuff on gekkoscience.com

Also 3 AntMiners at stock clock would require, if the measurements I did a while ago were accurate, about 960W. If you have a 1000W supply that you trust (like perhaps a DPS-800GBA on 220V power) or greater, it's probably possible. Two of these D750 boards load-balanced would supply 4 Ants at stock or 3 Ants overclocked without any issue. Might do 4 Ants overclocked, I don't recommend it but it has been done.

[precrime3]

FPS-800 GBA, you have a interface board for this? Don't know how to solder, and if not, I'll just buy a 1000w pus, or two 750ws load balanced (as you suggested, even though I don't know how to do that either :/) which ever is cheaper. Also where from your website do I buy? What methods do you accept? PM me with some info if you have time

[sidehack]

I don't have any DPS-800 boards for sale, we're still prototyping. All we have right now are the D750 boards. Load-balancing is pretty easy once you understand what's involved. Load-balancing and master on/off can be done with two power wires and two signal wires.

https://bitcointalk.org/index.php?topic=503423.0