From an accuracy perspective, higher firing rates are generally better, since they allow both faster and higher resolution transmission of information.
From a biological modelling perspective, you’ll want to choose firing rates that reflect whatever neural system you’re modelling. In biology, there’s an energy cost for every spike, so they need to balance accuracy and energy.
For neuromorphic computing, there are also costs associated with each spike. The exact costs depend on the architecture and network, including things like the amount of fan-out (how many different places does a spike need to be routed to), and how many spikes the routing fabric can handle.