0
$\begingroup$

I want to do some GPU computing with an NVIDIA card, and am deciding between having a GTX 960 with a 2GB or 4GB ram. Which one should I take? How much difference would these make in terms of the batch size I can use for mini batch gradient descent? Would this difference be significant?

Thank you for answers.

$\endgroup$
  • $\begingroup$ Just get one with 4GB - there's a price difference of 10EUR. $\endgroup$ – stmax May 19 '16 at 14:43
1
$\begingroup$

Generally memory cost increases linearly while mini-batch size increases. If batch size 32 costs you 2GB memory, then batch size 64 will cost you 4GB memory.

In practice, deep learning framework(for instance, keras) will load the entire full-batch into your memory, thus you cannot observe apparent memory increment or decrement during mini-batch tuning.

References

https://github.com/fchollet/keras/issues/146

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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