I tried to do mixed precision learning with fastai v1 on camvid (as shown by jeremy on the lesson 3 or part1v3).
Here’s the notebook: https://goo.gl/9BHyze
I ran this on p100 GPU on GCP and it works as expected.
This GPU doesn’t have speed optimizations for mixed precision learning.
But Mixed precision training thread shows that Volta series GPUs have built in speed optimizations for mixed precision.
So, I tried to run the above example with v100 GPU on GCP.
Then the kernel died. (I tried many times)
Anyone knows what’s going on here.