TPU is slower than GPU for convolutional networks #6022
Unanswered
RahulBhalley
asked this question in
code help: CV
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi all!
I trained a ResNet-152 based image classifier on CIFAR 10 dataset (batch size of 64). Training on single NVIDIA Tesla T4 GPU finishes in 2 minutes and 12 seconds whereas on 8 TPU cores, it takes 3 minutes and 58 seconds.
Is TPU slower than GPU for convolution network? And are only fully connected networks good for training on TPU?
Beta Was this translation helpful? Give feedback.
All reactions