Topic: Cairnsmore1 - Quad XC6SLX150 Board - page 83. (Read 286370 times)

To be fair, thats how all high efficiency PSUs work now. AC->DC 12v, DC->DC 12v->whatever for everything else.

That would be a dream ... watercool it too!

Well, at this point, it's hardly worth it to have 20ghps that uses ~5-7kw of power and instantly turns it into heat.... What I would love to see most is a single-slot PCIe card with a bunch of some kind of asic on it.

Does that mean you're scrapping the awesome rig?

Sooo... I have this PSU, and the wiring to handle it. Let see, 2360 watts - that should get me 39 boards, if they pull 60 watts each. Am I doin' it right?



Hey, I actually have 2 of them! And I have this too, for power distribution:



That brute will run 6 of those Dells, so 6 x 39 = 234 FPGAs.
234 x $640 = $149,760.00. A mere pittance.  

Most enterprise PSUs that I have seen are 12vdc bulk powersupplies that then feed an in-chassis 5vdc and 3.3vdc power supply that steps down 12vdc. Then again your idea of enterprise and mine may differ. To me enterprise = rack mount gear. There might be a market for an adapter that could accept a rack-mount PSU and output a metric shit ton of PCIe 6-pin connectors. Problem with that though is you have to deal with 40mm fans..

Its not hissy fit when there are other enterprise PSU that we can use. The 4-pin molex is still very popular outside of PC.

Here's a good overview about various connectors: http://www.modding-faq.de/index.php?artid=200
It's in german but you don't need to understand the text for it to be useful.

Thanks for checking, and that reinforces the fact that even if the connector fits it shouldn't be used that way.

It is actually the opposite of what you have listed rjk. PCIe has the 12v lines on non-latch side and the CPU4+4/EPS12v has the 12v lines on the latch side.

EDIT: Also, for the 8pin/6+2 PCIe cable the only addition from your current design is the '+2' which is only 2 grounds so not really helpful I think.

I believe it is correct that the EPS12V connectors have the positive and the negative swapped relative to a 6/8-pin PCIe connector, so be careful. Even if it fits, you might end up with reverse polarity and that would be bad.

So the 6-pin goes like this (with the connector latch on top):


And I believe the ESP12V goes like this (also with the connector latch on top):


But I don't have one next to me and so I can't check.

[/quote]

Most powersupplies can be maxed out on PCIe and +12v EPS pins. I don't know why you're throwing a hissy fit over obsolete and bulky 4-pin molex connectors. The better question is why aren't the 8-pin EPS connectors being utilized. 
[/quote]

The EPS certain would have ben nice and we could have got them in on a bigger board. It is a design compromise on the current size and we can always do another one if it looks the right thing to do. We could have shoehorned in more connectors but that might have meant more copper loss because we would have had to thin down the conduction paths. So as i say it is a compromise.

We did look at putting in the 4x2 PCIE as well but could not find the part number for the special keying it has and the Molex website doesn't help very much. Ironically the parts we got samples of could take take the EPS and had the right keying for that. If anyone has the 4x2 board side connector part number let me know and we can look at it for next time. I don't think it makes much difference on power handling they way they do it but might be easier for plugging in the 3x2+2x1 connectors currently common on ATX PSUs.

How this board can be used will be very dependent on how the host ATX PSU is configured and there are lots of variations there to deal with.

Liking the power board. When it is available will it be possible to just add those to our shipments if we request/pay for them?


Most powersupplies can be maxed out on PCIe and +12v EPS pins. I don't know why you're throwing a hissy fit over obsolete and bulky 4-pin molex connectors. The better question is why aren't the 8-pin EPS connectors being utilized. 

Wow! That sounds like enterpoints target about the bitstream is to bring the board on it's limit!
When I take the 800Mh/s at 40W as a standard for a quad board, that would mean nearly 1200Mh/s (I know it's only on my calculator ) at 60W! And I'm sure your team will make it. 

Yes we didn't use every last wire but there can be difficulties here as some ATX power supplies have seperate regulation stages and combining them together can mean that they fight and one regulator stage can be very stressed. So for the loss of 2 or 3 wires we left out the molex. That's only equivalent of one PCIE and we do have six of them. What would have been more useful to use was the 12V connectors for motherboards. In the end we went for a practical layout. I would have liked to have 8 PCIE but space didn't allow. However 2 PDBs can be used with one ATX in that case.

We didn't connect the PCIE together because of the fight issue so each of these has 3 Pheonix connectors on it and there are more or less independent.

Of course the remaining Molex can be used directly if you really want every last bit of power out of the PSU. Making simple adaptors for the motherboard 12V is also possible. The PDB gives you the easy way to use use the ATX PSU. There are a bunch of things we could have done and this was just the balance we choose.

If you do use the extra strings for any of this you do need to check individual output, or group, limits from the ATX PSU. That does go for the PCIE as well but much less of a chance of an issue here. We will try and build up a list of configurations we have checked or tested with a given PSU.

This product isn't just designed just for this week. It is designed for the limit of the Cairnsmore1 limit which is about 60W or 5A at 12V.

40 watt when use Icarus 380MH/s bitstream x 2.

I would say, the board can provide 12A per fpga, if there will be a nice bitstream then count with 12A, the boards limit.
Thats a maximum of 14.4 watt per fpga = 57.6 watt per board (running on it's limit)
18 boards at 57.6w = 1036.8 watt or 86.4A on the 12V.

It's better to have headroom like this if a nice bitstream will come and while there is no such a bitstream, your PSU will run at a nice efficiency. 

These are icarus based boards so 18 boards at 40w = 720w

There's plenty of headroom with 6 pci connectors.

Currently your board only limits 24pin and PCI-e connectors, i want to know why? Every PSU still has 2-3 molex 4-pin wires (i'm counting only wires because other connectors share the same wire are useless). Why not maximizing the available wires of a PSU? to save few bucks cents on the connector?

Shouldn't you be able to run 6 of these off 1 pcie connector?

Female pcie6--->6 molex

How much power are these using?

Actuallly I was looking at the post before. Yes that's very much the idea that you can do more or less what you want with the wiring. It's always difficult to get PCIE leads to wire up nicely in a rig always the wrong length and it's hard to split them nicely. Hopefully this board will be of use here. I imagine it will get used for other non-Cairnsmore uses as well. ATX PSU are hard to beat in efficiency terms and high power at a reasonable cost.

The plan is that this is an initial version and we will follow up with one supporting Ethernet as well. That enhanced one won't be done for 2-3 months yet. Depends on how busy we are,