May I know why it’s not enabled by default? What issues could it bring to the application?
Hi Keren,
Collection of latency timestamps i.e. ‘submitted’ and ‘queued’ timestamps adds overhead. Collection performance is improved in CUDA 11.3, but since not all users would be interested in latency timestamps, these are not enabled by default. There is no other issue which we are aware of.
Thanks for the answer. I’ll test and see if it brings significant overhead.