0

I was wondering what was the best manner to benchmark a GPU with CUDA over a specific application to obtain an accurate plot of GFLOPs/Watt.

So I would like to run a Kernel and get FLOP/Watt data to plot, and how I could do it efficiently.

Thanks

Anoracx
  • 438
  • 7
  • 24
  • You'll need to know how many FLOPs your application uses. Since you are referring to a specific application, nobody can help you with that unless you give some details about it. I don't know of any way to get the Watts consumed except by some electronic method like using a [Watts up meter](https://www.wattsupmeters.com/secure/index.php) on the system, or something even more complicated. There is power draw data reported by the nvidia-smi tool (for some GPUs), but it's probably too coarse a measurement for this. So this isn't really a programming question. – Robert Crovella Apr 24 '13 at 15:02
  • I guess you could compute/benchmark the GFLOPS and measure the energy consumption with something like [`PowerTOP`](https://01.org/powertop/). – BenC Apr 24 '13 at 15:06
  • So I guess the right question is first of to measure Flops for a specific Kernel ! – Anoracx Apr 24 '13 at 16:11

0 Answers0