An efficient hardware-aware neural architecture search is crucial for automating the creation of network architectures that are optimized for resource-limited platforms. However, challenges arise owing to inaccuracies in key hardware performance metrics, notably in latency estimation. This study introduces a composite loss-based complexity-driven latency predictor, which is an innovative approach that achieves remarkable evaluation accuracy with limited training data. This reveals a robust correlation between the layer-based complexity features and network inference latency. This groundbreaking insight leverages these complex features as network architecture encodings for latency predictors, substantially enhancing the precision of latency assessments. In addition, a composite loss function is proposed that seamlessly integrates ranking and absolute performance losses. This novel approach addresses the limitations of rank-based loss methods, which often lack broader context. Incorporating a global perspective through absolute performance metrics significantly improves the generalization capabilities of the predictor across various benchmarks. Experimental results on the NAS-Bench-201, NAS-Bench-101, and MobileNetV3 benchmarks underscore the effectiveness of the predictor. For instance, in the NAS-Bench-201 evaluation, the predictor demonstrates a notable increase in Kendall's tau correlation, from 0.738 to 0.9733. These findings highlight the enhanced accuracy of the proposed approach with far-reaching implications for optimizing network structures on resource-limited platforms.