2

Looking at watch -n 0.1 nvidia-smi I sometimes see that my GPUs' power consumption goes above the GPUs' official limit. Does that mean that GPU may actually go above their specified power consumption, or is the Watt measure noisy?

In the screenshot below, the 4th GTX Titan GPU is using 259W, while the maximum power consumption is 250W according to nvidia-smi in Nvidia specifications.

enter image description here

Typically it doesn't go over 250W for more than a few hundred milliseconds.

I use 4 GTX Titan GPUs on Ubuntu 14.04.4 LTS x64.

Video:

enter image description here

4
  • 1
    Do you have another application for monitoring the wattage? If you do, see what you get on that application? I wouldn't think that a few extra watts (maybe no more than 20 watts over its specs, especially for such sort duration's) are incorrect. Side note - you must have some serious video rendering or gaming with all that firepower!
    – Tim G.
    Commented Jul 31, 2016 at 2:23
  • @TimmyJim Thanks, I don't have another application for monitoring the wattage. Any suggestion? Commented Jul 31, 2016 at 2:24
  • 1
    Can a GPU draw more power then it is suppose to and/or the PCI-E specification allows, sure it can, AMD's current top of the line GPU does exactly. I should say it did before a software patch was issued.
    – Ramhound
    Commented Jul 31, 2016 at 2:26
  • 1
    It looks like the entirely expected result of voltage transients to me which ultimately are entirely expected observations given the physics of electrical circuits with varying loads. A detailed discussion of why this is an intrinsic property of electrical circuits would be more suited to physics.stackexchange.com but ultimately the speed of light is finite and the circuit has finite dimensions.
    – MttJocy
    Commented Jul 31, 2016 at 4:33

1 Answer 1

4

Answering this as an engineer more than as a superuser, yes, many devices can exceed their official maximums for brief periods, and frequently do, which is why we have such as thing as a slow blow fuse. As long as the condition is transient and doesn't happen 'too much', the fuse is content to allow the overage, because it got a chance to cool. If the overage is too long, or too frequent, then even a slow blow fuse will pop.

As this relates to you, the rating on the card is not an absolute theoretical maximum, it's an official average 'guess' based on a lot of calculation and observation of use cases. You, my friend, seem to have a use case that falls a bit outside these models. I'd be interested in knowing what you have happening on these cards. Bitcoin? Rendering? Running the entire control room?

As I watch the gif, I see the GPUs spend most of their time well below 250, and that at least in the data provided, only that last one, GPU4 is going over. You may just have a marginal GPU, or it might be doing more work than the others. The fan indicators dot seem to indicate a heat problem, and so unless this starts to be a problem, and as long as it remains consistent over time, I'd be comfortable calling this a curiosity and little else.

2
  • Thanks! I use the GPUs to train some artificial neural networks. Where do you see that the fan indicator dot seem to indicate a heat problem? (I'm not sure what you refer as "fan indicator dot") Commented Aug 2, 2016 at 15:14
  • To the left side of the big, the GPU fans are all in the 30s and 40s. If there was a heat problem, I'd expect they would be spinning their little hearts out, but these are fairly mellow.
    – root1657
    Commented Aug 2, 2016 at 16:33

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .