1

I made my simple deep learning model. I have only cpu on my desktop now. BUT I have to estimate the training time of the model in gpu setting. I don't need exactly accurate training time estimation. Approximate time is okay for me. Is there any method or metric for this? One of my thought is comparing GFLOPs, but it is also non-sense when I calculated..

ex. cpu: intel e5-2686 v4, gpu: nvidia k80

주은혜
  • 97
  • 1
  • 9
  • in short - no, it is not possible. GPUs are not "faster CPUs", these are different architectures, which will speed up specific operations 20x, while do nothing to others; memory copying can actually make things even slower. There are some rule of thumbs, but no "proper" way of doing that. – lejlot Nov 24 '17 at 22:01

1 Answers1

1

You can use a very naive approach of estimating the speed up by approximating the degree of parallelism in your data set (for example if it is ridiculously parallel this might be equal to the number of dimensions in your data-set), assuming best case implementation of concurrency with zero overhead of actually transferring the data, the math then becomes simple:

speedup=# of physical working items on GPU (usually # of CUDA cores in typical nvidia implementation) divided by # of physical working items on CPU (# of physical cores).

There are also differences which are not accounted for that have to do with different hardware architectures (ISR,design,etc.) These kind of calculations will vary greatly from real-life performance depending on your estimation of the models parallelism, implementation of concurrency and hardware.

yotabyte
  • 178
  • 1
  • 12