Yeah I hear it’s common practice now to avoid synchronizing GPU training kernels in order to speed things up, and it has positive regularization benefits and little downside.