Replicas themselves occupy already scarce device storage resources, and on extremely resource-constrained TinyML devices, whether replicas can fit and how many replicas can fit may be a more prior question than latency optimization.