The growing use of always-on intelligent edge devices has increased the need for energy-efficient on-device machine learning. Traditional digital microcontrollers running quantized CNNs cannot reconcile energy resource needs with inference performance. This work investigates the hypothesis that on-chip integration of event-driven SNNs with analog-mixed-signal neuromorphic hardware reduces per-inference energy consumption by at least 50% compared to digital MCU baselines, maintaining accuracy within ±2% points. Baseline CNN models are designed and converted to SNNs and deployed on both ARM Cortex-M4 digital MCUs and commercial analog neuromorphic chips. Comprehensive benchmarking with public datasets demonstrates that the analog SNN paradigm achieves substantial energy savings, submillisecond latency, and model accuracy that is on par with digital counterparts. The findings underline the emerging prospect for scalable, privacy-preserving, and continuous TinyML inference in battery-powered and resource-constrained settings. This work evolves TinyML toward more sustainable and pervasive applications and lays the foundation for further research in the area of on-device learning and ultra-low-power AI architectures.