can my power supply support two NVIDIA GPUs?
Anonymous
6 Nov 2013 19:32:09 UTC
Topic 197253
(moderation:
I am currently running Ubuntu 12.04 with NVIDIA drivers 331.17 in support of a GTX 650 Ti. The power supply is rated at 750 watts.
Specs for the 650 Ti say it requires a 400 watt power supply and that it draws 110 watts.
I would like to add a NVIDIA GTX 770. It's specs say it requires a 600 watt power supply and that it draws 250 watts.
I don't know if you add the watts required by both GPUs to determine if the total is less then the rated power supply or if there is some other "formula". Do you think that my 750 watt power supply is adequate?
can my power supply support two NVIDIA GPUs?
)
Well assuming that you are not running other devices and the PSU is of good quality, i think you should be ok.
I run 2 gtx460s on a Corsair TX-650 under Ubuntu. It is rare that this in total draws over 400W at the wall, typically under 350W. I suspect some other applications could push that up, based on the fan speed are about 60% on the GPUs.
Next build, i´d like to find a PSU which had some form of inbuilt monitor for temperature, voltages power in / out.
Indeed, if it's a good
)
Indeed, if it's a good quality PSU, it should work without issues and still have plenty of headroom.
The Graphic Card manufacturers recommendations have to take all those cheap, low quality PSUs into account, hence their recommended PSU sizes for any given card (especially high end cards) are usually extremely conservative.
I'm running 3 GPUs (HIS HD7850) rated @130W peak each plus a 95W TDP CPU on a CoolerMaster B500 500W PSU.
Actual peak power draw is 377W (full sustained load), which puts the PSU into a good efficiency area.
That's something cheap PSUs can't handle (for comparison, a no-name 500W PSU probably wouldn't even allow for system power-on).
For comparison, the manufacturer HIS states a recommended 500W PSU for one (1) of these cards already. It I pushed the cards into true peak power consumption (i.e. FurBench or alike tools), I probably would be really pushing it with my 500W PSU.
Quality brands are something like Enermax, Seasonic, Chieftec, Cooler Master, beQuiet!, Corsair, XFX etc. (personal preferences may vary amongst experienced users)
What might be more of an issue is the excellent cooling that your cards will require. Running fast GPUs in typical, hardly cooled midi tower cases often means trouble or a quick death for the hardware.
Excellent cooling (case fans) is the Alpha and Omega of GPU crunching...
Approximately a year ago, I
)
Approximately a year ago, I ran two GTX 680 cards and an Intel 3930K CPU via a Seasonic X-750 PSU. If I remember correctly, power draw was around 450w with the system dedicated to Einstein. The GTX 770 has similar specifications to the GTX 680.
With a decent quality PSU, I do not think you should have any issues running dual cards with this project. I find that it is worthwhile to invest in a Kill-a-Watt device to be able to see what the power draw is at the wall.
agentb: RE: Next
)
agentb:
I know exactly what you mean. Here is a horror story about what happened today. On the machine I wish to add another GPU to I decided to take a look at the power supply through the side grate. It is a tall tower and it easy to peek inside. To my horror, when I looked at the PSU the fan was NOT turning. I immediate removed the side panel and to my surprise a small cable had somehow dislodged, fallen through the fan grill and jammed the fan. I removed and secured the offending cable. Surprisingly enough the fan started turning. I thought for sure it would have burned out the fan's motor. There was a large surge of heat from the PSU - a Cooler Master. (no kidding). All of my linux boxes that have GPUs run a script that acquires the GPU temperature at a set periodicity and sounds an alarm if an arbitrary temp is violated. Also email is sent to my cell. If this PSU had had a temp sensor and a way to acquire it I would do the same. Sometimes things stop turning and when that happens you need to know.
falconfly:
No doubt that heat is the enemy. I live in quite a warm climate in central Florida and we are now entering the cool season. When I had the box built it was in a huge tower with fans over every hole it had. Of course this only matters if the fans are turning (see above). It has run quite cool and now the cooler weather is here I can push it a bit harder.
I appreciate both your responses. Running two GPUs seems like a viable option. Its really about available $s. So let me ask another question. Do either of you (or anyone else) think that a GTX 770 is overkill for E&H and that I could accomplish the same goal with another GTX 650 (smaller $s). Any thoughts gladly accepted.
RE: Approximately a year
)
I am not familiar with a "Kill-a-Watt" device. What is its function? Does it shutdown power at the plug if a predefined load is violated?
RE: RE: Approximately a
)
No, it's a measuring device. http://en.wikipedia.org/wiki/Kill_A_Watt
RE: I am not familiar with
)
It's a device to measure the power usage of whatever device you plug into it, instead of connection the power cord into the wall socket you plug the Kill-a-Watt into the wall and then the device into the Kill-a-Watt, the power used by the device is then displayed.
http://en.wikipedia.org/wiki/Kill_A_Watt
I use one of my Ultra X4
)
I use one of my Ultra X4 Modular 850-Watt PSU's to run one desktop with a 660Ti SSC and run the 650Ti OC card on another desktop using the same PSU and it works great.
The 770 still costs a bit too much to replace any of mine so I haven't got to test one yet but I like the 660Ti 2GB
The 650Ti 2GB on sale can be a good choice but I found that the speed can be quite different with them depending on which system you add them to.
I have 2 650Ti's that are OC'd and 2 550Ti OC'd that complete tasks at different times because of the different systems.
One 650Ti on my older 3-core completes a pair of tasks in 4hr 14min and the other in a quad-core takes 5hr 45mins to run the same pair of tasks here.
Basically the same thing happens with the two 550Ti's run the same way.
So one of my 650Ti's runs almost the same speed as the 660Ti does.
I usually run the newest drivers for this (I only use all of mine for running these tasks here)
You can take a look at mine if you like.
RE: RE: I am not familiar
)
I will buy one of these to determine the current load. Just happens that they are on sale at smarthome.com
MAGIC:
I did look at your computers. One of yours has about the same stats as one of mine. Interesting variation across different boxes for the same card.
While I am a dedicated user
)
While I am a dedicated user of consumer-grade power meters, there is a basic problem in using them to assess power supply adequacy: they don't sample at a fast enough rate to catch peaks in consumption.
Oddly enough, the power consumption monitoring function in my UPS appears to sample at an appreciably higher rate than any of my standalone power meters, which got me to noticing that the INPUT to the PC is quite a bit more time-varying than I would have guessed when the main PC load is a GTX660 running a couple of Einstein Perseus Arm Survey jobs.
As the input variation is substantially damped down relative to the PSU output variation by PSU capacitors, I can only speculate on how big the variation of power consumption by the GPU is at a time scale of milliseconds and tens of milliseconds, but suspect it to be quite large.
I'm an electrical engineer, and I used to scoff at the routine recommendation of gamers, BOINC GPU nerds, and the actual manufacturers of GPU cards to use what seemed to me excessive margin on installed power supply capacity to average demand, but I was assuming much less time-variation in demand than I now think to be the case.
So my current guess is that the seemingly excessive margin often advocated is not there to account for "all those cheap, low quality PSUs" but rather for time-variation.
And I acted on this--replacing 420W PSUs in my two GPU hosts which serve BOINC with Seasonic X650s. I think I have seem an improvement in reliability since the upgrade. The higher power consuming of the two hosts draws an average 218 watts at the wall during normal crunching.