Networked control systems (NCSs) integrate multiple engineering domains - control, communication, and computation - whose interactions must be jointly considered during system design. Models of the plant and the communication network can be developed with varying degrees of prediction accuracy and computational cost, yet existing model selection approaches typically operate within a single domain without accounting for cross-domain resource trade-offs. This paper proposes a systematic optimization framework that jointly minimizes prediction error cost and computation cost across all model domains, subject to a hardware-aware timing constraint derived from empirical characterization of the target platform. The framework categorizes NCS-relevant information into dynamic model domains, which provide selectable candidate models, and boundary domains, which supply constraints and empirical parameters. A discrete optimization procedure identifies the optimal model combination and inference configuration for a given use case. The approach is evaluated on a PMDC motor-actuated mass-spring chain controlled over an IEEE 802.11ax WLAN, with three plant models and four communication network models of increasing fidelity deployed on a Raspberry Pi 5 with Hailo-8L AI accelerator.