Neuromorphic Computing and Brain-Inspired AI Architectures for Energy-Efficient Intelligent Systems: A Review of Spiking Models, Hardware Platforms, and Edge Applications
Abstract
ABSTRACT
Artificial intelligence workloads are increasingly limited not by arithmetic capability but by the energy cost of moving data between separated memory and processing units. Neuromorphic computing addresses this limit by emulating the brain’s event-driven, massively parallel, memory-centric organisation through spiking neural networks (SNNs) and co-located memory and compute. This review synthesises landmark peer-reviewed work on spiking neuron models, learning algorithms, digital and analogue neuromorphic hardware, and edge applications. Five hardware platforms are compared within a common framework, and a first-order analytical model is derived to identify when spike-based computation saves energy relative to dense artificial neural networks (ANNs). Using published 45 nm operation energies, the model gives a break-even of about 5.1 spike events per synapse per inference, with roughly tenfold savings at 0.5 events. Reported platforms consume about 24–26 pJ per synaptic event, whereas a 32-bit off-chip memory access costs about 640 pJ. Training maturity, benchmarking and programming tools remain the principal barriers to adoption.
KEYWORDS: Neuromorphic computing, Brain-inspired AI, Spiking neural networks, Energy efficiency, In-memory computing, Event-based sensing, Edge intelligence, Memristor, Loihi, TrueNorth
Artificial intelligence workloads are increasingly limited not by arithmetic capability but by the energy cost of moving data between separated memory and processing units. Neuromorphic computing addresses this limit by emulating the brain’s event-driven, massively parallel, memory-centric organisation through spiking neural networks (SNNs) and co-located memory and compute. This review synthesises landmark peer-reviewed work on spiking neuron models, learning algorithms, digital and analogue neuromorphic hardware, and edge applications. Five hardware platforms are compared within a common framework, and a first-order analytical model is derived to identify when spike-based computation saves energy relative to dense artificial neural networks (ANNs). Using published 45 nm operation energies, the model gives a break-even of about 5.1 spike events per synapse per inference, with roughly tenfold savings at 0.5 events. Reported platforms consume about 24–26 pJ per synaptic event, whereas a 32-bit off-chip memory access costs about 640 pJ. Training maturity, benchmarking and programming tools remain the principal barriers to adoption.
KEYWORDS: Neuromorphic computing, Brain-inspired AI, Spiking neural networks, Energy efficiency, In-memory computing, Event-based sensing, Edge intelligence, Memristor, Loihi, TrueNorth