mersenneforum.org Can a Micro-ATX support a GPU as well as an ATX MB?
 Register FAQ Search Today's Posts Mark Forums Read

 2022-06-06, 13:44 #1 EdH     "Ed Hall" Dec 2009 Adirondack Mtns 5,261 Posts Can a Micro-ATX support a GPU as well as an ATX MB? I'm looking for a motherboard to support a Tesla M40 and use a E5-2630 (since I have some). I'm seeing lots of motherboards that all have a heat sink component right under where the double-wide card will sit. I have a current motherboard with that exact placement and it fails if I don't have a fan on it, so I can't use the GPU in that machine. I have found some Micro-ATX boards that have the PCIe connector on the outer edge, such that there would be no heat sensitive component under the card. I know most of the card's power is pulled through the 8-pin PS connection(s), but is there a substantial draw via the connector, which a Micro-ATX wouldn't supply long term?
 2022-06-06, 13:59 #2 M344587487     "Composite as Heck" Oct 2017 3·311 Posts Power wise it should, the power draw from a PCIe connector is meant to be up to 75W per the spec (some older cards exceeded that but should be irrelevant), any non-bottom-of-barrel board should be fine. The main thing to watch out for with PCIe connectors is that even if they are physically x16, often only the top one is electrically x16. If PCIe bandwidth is an important factor for your workload you should check the manual or markings on the board to make sure whichever slot you use is actually x16. Unless you are omitting case fans entirely I wouldn't be too worried about surface-mount components on the motherboard even if they do have a heatsink, even light airflow should be enough but you should direct it to where you think it's important anyway. I'm surprised matx boards exist for an old server architecture. I'm surprised it's worth using a 32nm CPU, you must have very cheap electricity to not be better off buying a cheap modern arch.
 2022-06-06, 14:53 #3 EdH     "Ed Hall" Dec 2009 Adirondack Mtns 526110 Posts Thanks! I'm kind of lazy and only like spending small amounts at once (or none, by using free/cheap old hardware). My electricity is lower than most, but on its way up, as everything is. I have a Supermicro board (dual Xeon) that I expected to use, but it has a heat sink component right where I described, that from the start gave me system hangs. I found that a fan on the heat sink kept it running, but when I added the GPU, it covered half the heat sink and with the fan only partially covering it, the motherboard started hanging up again. I also have two Z620s that should work, but they only have dual 6-pin PCIe connectors on the power supply. (The GPU needs two 8-pin.) I've thought of jumpering the other two pins, but if I do that inadequately, I'll melt the connectors. Edit: I suppose if you could offer a suggestion, I should look at something newer. . . Last fiddled with by EdH on 2022-06-06 at 14:56
 2022-06-07, 11:23 #4 M344587487     "Composite as Heck" Oct 2017 3·311 Posts Using 6 pin to supply 8 pin isn't ideal but it should work without jumping, I've done 6 pin in 8 pin and even no 6 pin at all in a 1x8+1x6, may be card dependent and pro cards may be more picky. 6 pin spec is 75W, 8 pin 150W, even if a 6 pin was at 150W that's still only ~12.5A split between at least 2 wires, aka there should be no physical danger from the wires (a particularly weak PSU potentially overloading the 12V rail 24/7 is the bigger concern IMO). You don't need to jumper the last two pins AFAIK, doing so just allows the use of a third 12V wire (which I've never seen not exist but you should double check if jumping). I've been out of the market for a while but know that Zen2/Zen3/intel-12th-gen are all reasonable choices new, Zen2/intel-10th-11th-gen might be reasonable 2nd hand if you can find them. intel 12th gen can be paired with DDR4 or DDR5, DDR5 being way more expensive. GPU's are much better at PRP so it might not make sense to get a good CPU+RAM combo depending on workload. intel have iGPU, the Ryzen's you'd want for compute do not (Ryzen APU's sacrifice cache to fit the iGPU), so there is more hassle getting a Ryzen system going to use a dGPU for display or going headless. Here's some rule of thumb builds using pcpartpicker: Ryzen 5600X 6c12t Zen3, ~$432, 2x8GB 3200, 700W 80+gold, https://pcpartpicker.com/list/W7fBv3 Potentially$80 can be saved downgrading CPU to 6c12t Zen2 4500, but that sounds like a low volume part. Zen2 2nd hand is the more likely way to do this. intel 12400, ~$448, 2x8GB 3200, 700W 80+gold, https://pcpartpicker.com/list/3w9zv3$30 can be saved going for the 12400F but that means no iGPU so a bit more hassle. If downgrading CPU fully to facilitate a mostly GPU workload there are cheaper options, you can skimp further on RAM (no need for 3200, can get away with 1x8 or 2x4), you can probably go for a cheaper PSU (maybe 80+bronze, maybe less wattage), etc. Alternatively you could get away with using the 2630's by trying to limit their power usage heavily as they are not the star of the show, but I think one of the problems with them is that there's less scope for power saving than with newer parts. Worth trying if you have a wattmeter and time.
 2022-06-07, 12:37 #5 EdH     "Ed Hall" Dec 2009 Adirondack Mtns 148D16 Posts Thanks for all the info! I'll have to study some more. My Tesla K20Xm refused to work with only a 6-pin PCIe connector plugged into the 8-pin on the board. The M40 has an adapter cable I could probably easily rewire. I have a 750W PS (CX750M) and lots of different cases. My current expectation is primarily Msieve LA and ECM for the M40. I'm running the K20Xm on a machine with a Core2 Duo CPU and 8G, so I'm thinking the 2630, with at least 16G should be quite a step up for the M40. OTOH, I'm running the ECM stage 2 residues on a second machine. I expect if I went the Ryzen route, it would take over the stage 2 residue work.
2022-06-07, 13:44   #6
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2·3·1,229 Posts

Quote:
 Originally Posted by M344587487 may be card dependent and pro cards may be more picky. ... GPU's are much better at PRP .
I can confirm that some GPUs fail to start if not connected at both aux power connectors.
Not sure how you mean better at. Some GPUs (including recent consumer oriented NVIDIA) are much better used at TF than other computation types (PRP, LLDC, P-1). Others are high performance at PRP etc and high performance/power. (Some NVIDIA server grade GPUs; AMD especially Radeon VII). Check SP/DP ratio and GPU memory bandwidth.

 2022-06-07, 15:21 #7 henryzz Just call me Henry     "David" Sep 2007 Liverpool (GMT/BST) 24·13·29 Posts Would it possible to power the gpu using a separate power supply? A second power supply can be used if you use something like https://www.amazon.com/XT-XINTE-Supp...5870470&sr=8-3 to start it along with the first. Assuming that the Z620's power supply has a 4-pin molex connector this should work I think.
2022-06-07, 17:37   #8
EdH

"Ed Hall"
Dec 2009

5,261 Posts

Quote:
 Originally Posted by henryzz Would it possible to power the gpu using a separate power supply? A second power supply can be used if you use something like https://www.amazon.com/XT-XINTE-Supp...5870470&sr=8-3 to start it along with the first. Assuming that the Z620's power supply has a 4-pin molex connector this should work I think.
Interesting that you should bring this up. The Z620 does have some PATA 4 pin molex connectors, that I contemplated routing into the PCIe connector in place of those two additional pins. The questionable part was whether that 12V was from the same bus in the PS. The PS in the Z620 is 800W, so from that perspective, it should be sufficient.

 Similar Threads Thread Thread Starter Forum Replies Last Post fivemack GPU Computing 1 2015-12-11 03:28 henryzz Hardware 1 2014-10-11 12:20 TheMawn GPU Computing 3 2014-07-13 02:31 jasong jasong 16 2012-04-05 13:01 Unregistered Information & Answers 5 2011-07-05 17:12

All times are UTC. The time now is 11:18.

Tue Feb 7 11:18:11 UTC 2023 up 173 days, 8:46, 1 user, load averages: 1.08, 1.22, 1.15