1

I am executing a CUDA kernel in my A100 GPU and I've realized that the power consumption at some points is higher than nvidia-smi given range:

enter image description here

The picture has been taken from nvtop.

Is it something that I should be worry about?

Improve this question
1
  • "burst" performance has, for at least the last 5 years, been a fluid thing in computing. Intel processors have "turbo boost" where they self-overclock depending on thermal limits and power capabilities of the chip itself. Those overclocks are also time and instantaneous temperature dependent so a "cold" chip will overclock higher and that clock speed will drop over time as the device warms up. I would be surprised if Nvidia do not have similar capabilities.
    – Mokubai ♦
    Dec 24, 2020 at 10:51

2 Answers 2

Reset to default
1

The only worrisome aspect is the temperature, which seems to be at an unimpressive 52°C. This doesn't make sense if the power draw is truly above the max.

So, take your pick. Either:

  • The power draw figure is false
  • The temperature reported is false
  • nvtop is not working correctly with your GPU.

I would suggest verifying the temperature using other applications. If they also report the same readings, then you don't need to worry. Check both CPU and GPU and motherboard.

Useful references:

Improve this answer
2
  • Thanks for your answer. I use nvtop for an easier monitorization. Data is the same that in nvidia-smi.
    – Bub Espinja
    Dec 24, 2020 at 10:03
  • There is also above a fourth option, that the card is reporting false values. In any case, I would suggest keeping an eye on temperature readings, but otherwise, as long as the computer doesn't crash, believe in those temperature readings.
    – harrymc
    Dec 24, 2020 at 10:07
1

The power draw of a GPU is uneven - it has spikes and lows. The specified power draw of a card is ment to be read as "rolling average over one second" during which time it can fluctuate over and belo that value - this is one of the reasons why PSU specs are recommended to be way over the sum of specified component power draw in a GPU-heavy rig.

nvidia-smi and friends report not the rolling average, but the momentary power draw, which can of course exceed the specified value. If you randomly sample your GPU power draw over a statistically relevant sample, you are very likely to see a value very close to specs.

Improve this answer

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.