I am working on IMDB sentiment analysis using embedding + CNN. Why is there a big difference in computation time, with GPU (~9 seconds per epoch) being faster than TPU (~17 seconds per epoch), even though TPUs are supposed to be more powerful?
2 Likes
This comment has very little context, however it may be worthwhile to experiment with higher sample sizes and see whether it makes a difference.
1 Like
Google Colab offers free access to both T4s and TPUs for experimentation and lets you test which one works best for your specific workload.
1 Like
I experienced comparable problems. Look at this.