Latency Budget Relaxation

A task was originally fully online, making recurrence or causal convolution preferable. If the deployment changes to offline batch scoring with the whole sequence available, which architecture family gains the most from that relaxation?