Memory use by caffe

30 views

Skip to first unread message

Artyom Beilis

unread,

Sep 8, 2021, 3:48:27 PM9/8/21

to Caffe Users

It isn't a secret that Caffe is far from being optimal in terms of memory use.

For example In can bring at most batch size 12 of Resnet50 on 8GB gpu while I can easily reach 32 with frameworks like pytorch or keras.

There are lots of memory duplications and there is no optimization of memory reuse of intermediate results in backward pass that can save ~50% of memory.

Is there any branch that attempted to fix the problem?

I'm working on integration of caffe+dlprimitives but I found it is very hard to compare to other frameworks due to poor memory optimization.