The proliferation of Internet of Things (IoT) devices has amplified botnet risks, while traditional network-based intrusion detection systems (IDSs) struggle under encrypted and/or sparse traffic. Power consumption offers an effective side channel for device-level detection. Yet, prior studies typically focus on a single model family (often a convolutional neural network (CNN)) and rarely assess generalization across devices or compare broader model classes. In this paper, we conduct unified benchmarking and comparison of classical (SVM and RF), deep (CNN, LSTM, and 1D Transformer), and hybrid (CNN + LSTM, CNN + Transformer, and CNN + RF) models on the CHASE’19 dataset and a newly curated three-class botnet dataset, using consistent preprocessing and evaluation across single- and cross-device settings, reporting both accuracy and efficiency (latency and throughput). Experimental results demonstrate that Random Forest achieves the highest single-device accuracy (99.43% on the Voice Assistant with Seed 42), while CNN + Transformer shows a strong accuracy–efficiency trade-off in cross-device scenarios (94.02% accuracy on the combined dataset at ∼60,000 samples/s when using the best-performing Seed 42). These results offer practical guidance for selecting models under accuracy, latency, and throughput constraints and establish a reproducible baseline for power-side-channel IDSs.