Sorry if this is the wrong forum for this question. My understanding was GPUs were more efficient for running neural nets, but someone recently suggested GPUs are only needed for the training phase. Once trained, it's actually more efficient to run them on CPUs.

Is this true?


closed as off-topic by David Richerby, Evil, D.W. Nov 8 '16 at 23:24

This question appears to be off-topic. The users who voted to close gave this specific reason:

If this question can be reworded to fit the rules in the help center, please edit the question.

  • 1
    $\begingroup$ I don't think questions at this level of implementation-dependence are on-topic, here. $\endgroup$ – David Richerby Nov 6 '16 at 10:32
  • $\begingroup$ @DavidRicherby ok sorry! any suggestions on better forums? $\endgroup$ – Crashalot Nov 6 '16 at 20:08
  • 1
    $\begingroup$ Might be on-topic at Stack Overflow? But check their help centre first. (And sorry for not just saying that in the first place instead of waiting for you to ask.) $\endgroup$ – David Richerby Nov 6 '16 at 21:15
  • $\begingroup$ @DavidRicherby no problem! obviously you were just trying to filter off-topic questions and had only pure intentions. didn't post on SO because the likelihood of it being off-topic is probably 80%. $\endgroup$ – Crashalot Nov 6 '16 at 21:42
  • $\begingroup$ I don't know their scope perfectly, but Computational Science may have the expertise. That said, I think there is a CS question in there: which model of computation fits which aspect of working with neural networks better? $\endgroup$ – Raphael Nov 8 '16 at 17:01

Running trained artificial neural network is repeated performing of a matrix with vector multiplication and performing the same nonlinear function for each element of a vector. If both operations could be parallelised and vectorised, it is what is GPU well suited for. But it strongly depends on the size of given vectors and matrixes which depend on the size of the neural network. For small nets, CPU could be faster than GPU but it would not matter anyway. For huge nets, GPU would be probably much better suited if algorithm is written right.


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