Training a neural network with a video card and a processor - what's the difference in memory consumption?

Question

Neural networks

Now I am learning to work with neural networks. I'm trying everything in Python 3.6 using Keras. For example, training the current neural network "ate" about 3GB of RAM. If you transfer the training of such a network to a video card, will it also require 3 GB? or more? or less?
The network itself is simple - about 150k of trainable parameters, the rest is a training set.
Asking because I don't have a discrete graphics card to try.

Answer 1

Don't you just want to put it in the cloud?

Answer 2

Depends on the number of parallel threads: reduce it - it will take up less RAM, but it will take longer to execute.
A discrete card - of course, calculates faster, so the memory consumption of the card will depend on the number of parallel threads, and their number will depend on the structure of the neural network itself.