FP16 runs slower than FP32 on V100?

[dustinjoe]

Hi, I am trying to run a little profiling measurement on Whisper on a Tesla V100 GPU.
The performance of FP32 when FP16=False seems to be faster than the default option of FP16=True.
Generally seems to be about 1/4 faster.
I attach my test codes and results here(please remove the .txt in filename). Not sure anyone could give some suggestions on this? Thanks.

test_whisper_time_batch0.sh.txt

whisper_profile_time_args0.py.txt

whisper_profile_results0.rtf.txt

[tnlin]

same there, can't figure out why