Neural networks in embedded systems confront computational efficiency and energy consumption challenges arising from neural network inference. Although existing approximate computing techniques can reduce energy consumption, they lack runtime-reconfigurable approximation levels, making it difficult to meet the precision requirements of diverse scenarios. In this paper, we propose a time-efficient methodology that maps runtime-reconfigurable approximate multipliers to weight intervals while respecting tight accuracy-loss thresholds. It jointly considers per-layer computational volume and weight-interval fault-tolerance in 8-bit quantized networks, addressing prior neglect of volume disparities and misjudged weight significance. Experiments on CIFAR-10, CIFAR100 and GTSRB show that our method achieves average energy savings of 21.82% (up to 24.7%) on ResNet and outperforms state-of-the-art methods by \({>}\) 20% on VGG. Configuration completes in 0.6 h on GPU, yielding a 7.5x speedup over prior works.