larger batch size increases training time?

[Houssembenmid]

greeting, I have tried to increase the batch size for semantic3D from 16 to 64 as I thought it will decrease the training time .However, I got the opposite: the training time was increased respectevly with the batch size increase. I am actually confused is there an explanation for this?

[aboulch]

Hello, Sorry for late response, is your question still valid ? Part of the computation is done on CPU (nearest neighbor search essentially), it could be due to cpu/memory lagging.