machine learning - Why does Tensorflow CNN use too much memory? -

i'm new deep learning , i'm using tensorflow train cnn image recognition. training images 128 pixels * 128 pixels * 3 channels. in network, there 3 conv layers, 3 maxpooling layers , 1 connected layers. have more 180,000 labeled images decided train 4000 images each batch. however, training process can not run on laptop because memory not enough tried use sever 64gb ram , 2 * e5 cpu train it. time works, costs more 40gb of memory. i'm confused images used training not high resolution images(only 128*128). why still costs memory(may batch size big....). normal? if normal, how people use gpu train neural networks, far know, gtx1080ti has 11gb memory, still not enough training network.

4000 sounds lot in 1 go. examples i've seen train few hundred in each batch. imagine images may getting loaded memory @ once, hence high memory usage.

can try training smaller batches? 1000, or 500, , see if memory usage drops?