Another way of thinking about this is whether you want to optimize your workload for throughput or latency. It's almost never a binary choice, though.