The documentation states:
Deterministic mode can have a performance impact, depending on your model.
My question is, what is meant by performance here. Processing speed or model quality (i.e. minimal loss)?
Performance relates to the processing speed in this case. E.g. if you disable cudnn.benchmark, your execution might get slower (depending on the model).
Thank you for the confirmation. I feel that this may be clarified in the documentation, but leaving it answered here is fine too.
It might be confusing others too, so feel free to create a GitHub issue so that it can be discussed (and fixed) there, too. 
1 Like