Tesla K40 + GeForce GTX 780 minimum recommended power

I was wondering what is the minimum recommended power for a combination of a GeForce class card and a Tesla K40 GPU? This is installed in a workstation, so the GeForce is required to get a display output.

I currently run this with a 600W PSU, which frankly surprises me, as the system boots and no insufficient power warnings are displayed in Windows. However, the system seems to be unstable at times, I was wondering if it could be caused by this. If I went to get a stronger one, what should be the rating?

There is no CUDA code running on the GeForce, it is really only used for display.

You can discover a lot just by googling around.

600W is probably not enough if you heavily load the system.

The K40c can use up to 235W at full load:

http://www.nvidia.com/content/PDF/kepler/Tesla-K40-Active-Board-Spec-BD-06949-001_v03.pdf
http://www.nvidia.com/content/PDF/kepler/Tesla-K40-PCIe-Passive-Board-Spec-BD-06902-001_v05.pdf

The GTX780 requires up to 250W at full load, the and recommended PSU is 600W for a system containing just a single GTX780:

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-780/specifications

(if you have a GTX780Ti the numbers are a bit higher).

At 485W total for the above, that leaves just 115W (in a 600W PSU) for everything else. Some intel CPUs by themselves can require 115W or more (again, when heavily loaded).

Insufficient power can lead to instability. You might be better off with an 850W PSU.

Naturally, you may wonder “how much power does an unloaded/lightly loaded GTX780 consume?”

I don’t know the answer to that. It seems like 850W could increase your confidence that power supply is not an issue in an instability observations. But it might be that 600W is enough if you don’t do much with the GTX780. That kind of data is not officially published as far as I know.

I concur with txbob’s recommendation. My own heuristic is to add up the maximum power ratings for all GPUs and CPUs, add 40W for motherboard and peripherals (HD or SSD, ethernet), plus 1W per GB of DDR3 system memory, then multiply the sum by 1.2 as a safety margin that protects against manufacturing tolerances, component aging etc.

In my experience, using a PSU with adequate reserves is not just a question of momentary system stability, it can also have an impact on PSU longevity. I burned out a few PSUs over the years working with CUDA by running them close to 100% load. Since heat from all components can also be an issue with a heavily loaded system (not to mention the cost of the electricity bill), as a personal choice, I usually look for efficient PSUs, e.g. 80 Plus Gold rated.

Thanks for the answers, what you say makes a lot of sense. I will definitely be getting a stronger PSU, just wanted some inputs first to not mess up :).