Neuromorphic edge intelligence for next-generation IoT systems: architecture, prototyping, and performance evaluation

As the Internet of Things (IoT) devices are rapidly increasing in number, the necessity of energy-efficient and low-latency data processing in the edge grows. Conventional cloud- and GPU-based solutions are not typically able to satisfy such demands, especially in resource-limited settings. Neuromorphic computing provides an alternative opportunity with event-based processing and spiking neural networks (SNNs) to provide real-time intelligence. This study fills that gap by creating, prototyping, and testing neuromorphic IoT devices that can be used in practice, in both a medical context and smart residence. This study gives the design, prototyping and evaluation of a neuromorphic IoT, based on the Intel Loihi 2 and Innatera Pulsar, with real-world sensors. System validation is done by field testing and large scale simulation. The prototyping used Intel Loihi 2 and Innatera Pulsar connected to real-world sensors. Afterward, field testing and scalability modeling is carried out using the NS-3 and OMNeT++. Benchmarks on a traditional GPU-based platform (NVIDIA Jetson Nano) show that latency can be improved by 87.3%, energy consumption by 89.7%, with a 1.2% increase in the accuracy. Experimental demonstrations show its practical applications, such as healthcare monitoring, abnormal physiological scores annunciated in less than 2 ms; smart houses, event-driven sensing had reduced consumption of needless energy by 28% compared to rule-based systems. Scaling simulations further validated the system performance across networks of up to 10,000 nodes. The average latency remained below 15 ms, demonstrating the capability of the proposed architecture for real-time large-scale IoT deployments. In addition, the packet delivery ratio (PDR) remained above 98% for up to 8,000 nodes and achieved 96% at 10,000 nodes, confirming reliable communication under increasing network density.